Upload
others
View
4
Download
1
Embed Size (px)
Citation preview
Version: v1_8
Date: 22 Aug. 2013
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 1/139
PROBA-V Image Quality Center (IQC) ATBD
(US-05)
Status information
Security classification: Project Confidential State: Final (for acceptance)
Current version number: v1_8 Date of first issue: 22/06/2009
Prepared by: Sindy Sterckx, Wouter Dierckx, Ils Reusen,
Stefan Livens, Stefan Adriaensen, Geert
Duhoux, Jan Biesemans
Date: 14/05/2013
Verified by: PROBA-V PA team Date: 22/05/2013
Approved by: Jan Dries Date: 27/05/2013
Copyright notice
This document may not be reproduced (even partially) or communicated to third parties without the written authorisation of VITO’s General Management.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 2/139
Circulation / distribution list
Name Address P/E Name Address P/E
PROBA-V team VITO E Karim Mellab ESA E
Juliette Kaysan ESA E
Joe Zender ESA E
(P = Paper copy, E = Electronic version)
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 3/139
Document change record
Version Date Description Affected sections
Editor(s)
0_1 29/10/2009 Initial (draft) version All Sindy Sterckx, Wouter Dierckx, Ils Reusen, Stefan Livens, Stefan Adriaensen, Geert Duhoux, Jan Biesemans
20/11/2009 PA modifications 2;3 Wouter Dierckx
1.0 10/02/2010 First release All sections Sindy Sterckx, Wouter Dierckx, Ils Reusen, Stefan Livens, Stefan Adriaensen, Geert Duhoux, Jan Biesemans
1.1 17/06/2010 General review and update to current status:
- sensor model is now with a fixed ADC gain and with nonlinearity correction as a first step
- masked edge pixels for dark current monitoring are no longer foreseen
- MTF algorithm description is compacted to the essential
3
4.1
4.3
4.9
Wouter Dierckx
22/06/2010 Change of solar irradiance file
Update Visio with radiometric calculation, in line with sensor model update
Use of SPS and SPOT-VGT data for validation clarified
Updated graphs in Section on Cross Sensor calibration
4.4.1.2
Figure 10
4.10
4.10.1
Sindy Sterckx
Stefan Livens
1.2 14/09/2010 Corrected formulae in linear weighted model
4.6.2.3.2 and 4.6.2.3.4
Stefan Livens
1.3 28/09/2009 Sun glint error budget updated
NIR reflectance threshold modified
Resolution LUT modified
Aerosol data replaced by temp data
Table 13
4.5.1.2.2.2
Table14
Figure 13
Sindy Sterckx
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 4/139
Version Date Description Affected sections
Editor(s)
1.4 28/04/2011 DCC cal : distance to coast removed; MODIS temp check removed
Absolute calibration deserts modified
Multi-temporal calibration deserts modified
Cross-sensor calibration deserts modified
Geometric calibration chapter deleleted, reference to ATBD geometry
4.5.2
4.4.4
4.5.3
4.10.1
Sindy Sterckx
Sindy Sterckx
Sindy Sterckx
Stefan Livens/Sindy Sterckx
1.5 16/09/2011 Rayleigh: site depended CHL added
cross-sensor deserts : outlier selection reference added
Camera-to-camera : Usage of mean relative difference instead of mean difference
Moon calibration: Under investigation removed
Multiangular Calibraiton Antarctica: added theoretical background
4.4.1.2.1.4
4.10.1.2.3
4.7.2.1
4.5.5
4.8
Sindy Sterckx
Sindy Sterckx
Sindy Sterckx
Sindy Sterckx
Stefan Livens
1.6 25/10/2011 Update of Defect Pixel algo after internal review
4.3.2 Wouter Dierckx
1.7 27/03/2013 Update Radiometric Sensor Model
Update thermal behaviour
4.2
4.3
4.7.2.2
4.8.3
4.2.1.1
Sindy Sterckx
Wouter Dierckx
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 5/139
Version Date Description Affected sections
Editor(s)
1.8 03/04/2013
26/04/2013
2/05/2013
13/05/2013
Correction of dark current relation formula
Specification of the units of the radiometric model coeff
Add AVG in inverse modeling
ADD normalization by IT for dark current
Correction of 2 equations: add normalization by integral over spectral response
IQC-RC straylight handling
Lunar Image Processing
Dark current LUT
Seasonal correction deleted
4.2.1.1
4.2.1
4.2.2
4.3.1.2
Eq. 23 en 30
4.5.5
4.11
4.2.1.1
4.4.2.2.8
Wouter Dierckx
Sindy Sterckx
Sindy Sterckx
Sindy Sterckx
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 6/139
TABLE OF CONTENTS
1. INTRODUCTION 10
1.1 Scope 10
1.2 Applicability 10
2. REFERENCES 11
2.1 Project reference documents 11
2.2 Other reference documents 11
3. TERMS, DEFINITIONS AND ABBREVIATED TERMS 18
4. RADIOMETRIC CALIBRATION 19
4.1 Introduction 19
4.2 Sensor Radiometric Model 22 4.2.1 Model overview 22 4.2.2 Inverse model 24 4.2.3 Operational dependencies 25
4.3 Radiometric Characterization 26 4.3.1 Dark Current 26 4.3.2 Detection of defective and degraded pixels 29 4.3.3 Linearity Check 33
4.4 Absolute radiometric calibration 34 4.4.1 Rayleigh calibration: oceans 34 4.4.2 Reflectance-based method 45 4.4.3 Radiance-based method: APEX underflights 53 4.4.4 Absolute calibration over stable deserts 61
4.5 Relative radiometric calibration 67 4.5.1 Interband: Sun glint 67 4.5.2 Interband: deep convective clouds 76 4.5.3 Multi-temporal Calibration over stable deserts 84 4.5.4 Multi-temporal Antarctica 88 4.5.5 Multi-temporal Moon 89
4.6 Statistical analysis on calibration coefficients 92 4.6.1 Introduction 92 4.6.2 Approach 93 4.6.3 Updating the operational coefficient 102 4.6.4 Dealing with season-dependent errors 104
4.7 Camera to camera calibration 105 4.7.1 Introduction 105 4.7.2 Algorithm 106
4.8 Multi-angular/Equalization 111 4.8.1 Introduction 111 4.8.2 Theoretical background 112 4.8.3 Algorithms 114
4.9 Performance Indicators 119 4.9.1 Noise analysis 119 4.9.2 MTF assessment 125 4.9.3 SNR calculations 126
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 7/139
4.10 Calibration Validation 128 4.10.1 Cross-sensor calibration over stable deserts 129
4.11 IQC-RC Straylight handling 136 4.11.1 Introduction 136 4.11.2 Impact on Calibration Methods and proposed actions 136 The largest impact is expected for those methods using high contrasting
calibration scenes such as : 136
5. GEOMETRIC CALIBRATION 138
Appendix 1 FLOWCHART CONVENTIONS 139
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 8/139
LIST OF TABLES
TABLE 1: PROJECT REFERENCE DOCUMENTS 11 TABLE 2: LITERATURE REFERENCES 17 TABLE 3: ABBREVIATIONS AND ACRONYMS 18 TABLE 4: PIXEL QUALITY CALCULATION 29 TABLE 5: THRESHOLDING VALUES FOR DEFECT DETECTION 32 TABLE 6: STABLE OLIGOTROPHIC OCEANIC SITES 39 TABLE 7: SUMMARY OF RAYLEIGH ABSOLUTE CALIBRATION ERROR BUDGET. 45 TABLE 8: CEOS CORE INSTRUMENTED IVOS SITES (LANDNET SITES). 46 TABLE 9: RADIOMETRY REFERENCE TEST SITE SELECTION CRITERION. 47 TABLE 10: ERROR SOURCES OF REFLECTANCE-BASED METHOD. 53 TABLE 11: APEX MAIN SPECIFICATIONS 54 TABLE 12: ERROR SOURCES FOR RADIANCE-BASED METHOD WITH REFERENCE TO NIST
STANDARDS. 61 TABLE 13: SELECTION CRITERIA AND REQUIREMENTS FOR SUITABLE DESERT SITES 61 TABLE 14: DESERT SITES USED FOR RADIOMETRIC CALIBRATION 62 TABLE 15: SEA WATER INDEX OF REFRACTION FOR THE PROBA-V BANDS 70 TABLE 16: SUMMARY OF SUN GLINT ABSOLUTE CALIBRATION ERROR BUDGET 75 TABLE 17: PROPERTIES OF THE CLOUDY ATMOSPHERE AND SURFACE FOR DCC LUTS 80 TABLE 18: CALIBRATION ERROR BUDGET FOR DCC 83 TABLE 19 : NOTATIONS 93
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 9/139
LIST OF FIGURES
FIGURE 1: CONCEPTUAL OVERVIEW 19 FIGURE 2: INITIAL QUALITY STATUS IN RED FOR 30 PIXELS. 28 FIGURE 3: DARK CURRENT DATA FOR VGT-MIR DATA (FROM 512 LINES) 28 FIGURE 4: MEASUREMENT OF LINEARITY, BY VARYING INTEGRATION TIME. 33 FIGURE 5: ILLUSTRATION OF LINEARITY DEGRADATION. 34 FIGURE 6: GASEOUS ABSORPTION IN PROBA-V BANDS 38 FIGURE 7: MONTHLY AVERAGED CHLOROPHYLL (IN MG M
-3 ) FOR EACH RAYLEIGH
CALIBRATION ZONE 40 FIGURE 8: OVERVIEW FLOWCHART RAYLEIGH CALIBRATION 43 FIGURE 9: FLOWCHART TO CALCULATE CALIBRATION COEFFICIENTS 52 FIGURE 10: FLOWCHART TO CALCULATE CALIBRATION COEFFICIENTS 59 FIGURE 11: FLOWCHART FOR APEX-BASED TOA RADIANCE FOR PROBA-V SPECTRAL BANDS 60 FIGURE 12: LOCATION OF DESERT SITES USED FOR RADIOMETRIC CALIBRATION, 62 FIGURE 13: AOT SEASONAL VARIATION 64 FIGURE 15: SUN GLINT GEOMETRY 69 FIGURE 16: OVERVIEW FLOWCHART DCC 74 FIGURE 17: SCIAMACHY TOA REFLECTANCE SPECTRA OVER DCC 76 FIGURE 18: DCC TOA REFLECTANCE FOR IN FUNCTION OF CLOUD OPTICAL DEPTH. 77 FIGURE 19: DCC TOA REFLECTANCE IN FUNCTION OF CLOUD EFFECTIVE RADIUS. 77 FIGURE 20: MIXING SCHEME ACCORDING TO BAUM ET AL (2005) 79 FIGURE 21: FLOWCHART DCC CALIBRATION 82 FIGURE 22: FLOWCHART MULTI-TEMPORAL CALIBRATION OVER DESERTS 85 FIGURE 23: EXAMPLE OF MULTI-TEMPORAL VARIATION (TAKEN FROM SPOT-VGT2) 88 FIGURE 24: LOCATION OF DOME C CALIBRATION SITE (75°S, 123°E) 89
FIGURE 25: CALCULATION OF BEST ESTIMATE OF kA 92
FIGURE 26: SPATIAL AVERAGING 95 FIGURE 27: RESULT OF LINEAR REGRESSION, WITH 1 AND 2 SIGMA CONFIDENCE INTERVALS 96 FIGURE 28: TIME WINDOW WITH LINEARLY INCREASING WEIGHTS 97 FIGURE 29: RESULT OF LINEAR REGRESSION, INCLUDING TEMPORAL WEIGHTING. 98 FIGURE 30: RESULT OF LINEAR REGRESSION, INCLUDING TEMPORAL WEIGHTING AND
INDIVIDUAL ACCURACIES 98 FIGURE 31: VALUES FOR WHICH THE DIFFERENCE IS JUST SIGNIFICANT AT THE GIVEN
CONFIDENCE LEVEL 103 FIGURE 32: OVERVIEW OF THE VIEWING GEOMETRY FOR ONE SCAN LINE. 106 FIGURE 33 CAMERA OVERLAP ZONE 107 FIGURE 34: REGRESSION OF TOA RADIANCES CAMERA 1 VERSUS CAMERA 2 108 FIGURE 35: REGRESSION OF THE DIFFERENCES IN TOA RADIANCES AGAINST THEIR MEAN 108 FIGURE 36: SENSITIVITY PROFILE DECOUPLED INTO HIGH AND LOW FREQUENCY
COMPONENTS 111 FIGURE 37: SURFACE PROFILE, TWO DETECTOR SENSITIVITY PROFILES AND THE COMBINED
RESULTS RESULTS 112 FIGURE 38: ENCLOSING RECTANGLE IN IMAGE COORDINATES FOR A RECTANGLE IN
GEOGRAPHIC COORDINATES 115 FIGURE 39: LINE SEGMENTS TO AVERAGE WITHIN THE RECTANGLE OF INTEREST 116 FIGURE 40: CROSS SENSOR CALIBRATION OVERVIEW 131 FIGURE 41: CONVENTIONS USED IN PROCESSING FLOW DIAGRAMS 139
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 10/139
1. INTRODUCTION
1.1 Scope
The present document is the Algorithm Theoretical Basis Document (ATBD) for the Image
Quality Center of the PROBA-V User Segment (PV-02) project), under contract between VITO
(supplier, Mol, Belgium,) and ESA (customer) as part of the PROBA-Vegetation (PROBA-V)
project.
In the project Document Requirement List (DRL) [PVDOC-005] it is referred to as US-5.
The document provides a detailed description of the different algorithms that compose the Image
Quality processing chains to monitor the sensor performance and to determine the parameters to
be supplied to the processing facility (PF).
1.2 Applicability
This document defines the Algorithms Theoretical Basis of the Image Quality Center.
It has to be delivered to the customer at US-PDR, US-CDR. This document might be subject to
change between the above mentioned review cycles and will reach a “final (accepted)” status at
US-CDR.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 11/139
2. REFERENCES
All applicable and reference documents for the PROBA-V PV02 project, either initiated by ESA
or the consortium, are listed in
[N77D7-PV02-PM-18-US-ApplicableAndReferenceDocumentsList].
2.1 Project reference documents
Project reference documents are listed in Table 1:
[PVDOC-005] PROBA-V User Segment: Document Requirement List
(DRL) and Document Requirement Description (DRD)
[PVDOC-623] Detailed Processing Model IQC
[PVDOC-615] User Segment Calibration Plan
[PVDOC-624] Detailed Processing Model PF
[PVDOC-611] Technical Note on Sun Glint
[PVDOC-034] PROBA-V Vegetation Instrument : Thermal Design and
Analysis Report
[PVDOC-068] PROBA-V SPS Use Cases
[PVDOC-647] PROBA-V TN on Radiometric calibration algorithms
prototype results
[PVDOC-981] Algorithm Theoretical Baseline Document Geometric
Calibration
[PV-TN-ESA-SY-0010] Proba-V straylight measurement analysis
[PVDOC‐190] PROBA-V straylight additional analysis II (PV-TN-ESA-
SY-0008)
[PVDOC‐655] PROBA-V TN: effect of straylight on Rayleigh Calibration
Table 1: Project reference documents
2.2 Other reference documents
Other reference docs (e.g. literature list) are listed below:
LIT1 Govaerts and Clerici, 2004 Govaerts, Y. and M. Clerici, 2004.
Evaluation of radiative transfer simulations
over bright desert calibration sites, IEEE
Transactions on Geoscience and remote
sensing 42(1), 176-187.
LIT2 Vermote et al., 1992 Vermote, E., R. Santer, P.Y. Deschamps, M.
Herman, In-flight calibration of large field-
of-view sensors at short wavelengths using
Rayleigh scattering, Int. J. of Remote
Sensing, 13, No18, 1992.
LIT3 Hagolle et al., 1999 Hagolle, O., P. Goloub, P-Y. Deschamps, H.
Cosnefroy , X. Briottet, T. Bailleul, J.-M.
Nicolas, F. Parol, B. Lafrance, M. Herman,
Results of POLDER In-Flight Calibration,
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 12/139
IEEE Transactions on Geoscience and
Remote Sensing, Vol. 37, No. 3 (1999)
LIT4 Fougnie et al., 2007 Fougnie, B., G. Bracco, B. Lafrance, C.
Ruffel, O. Hagolle, C. Tinel, PARASOL in-
flight calibration and performance, Applied
Optics, vol. 46, No. 22 (2007)
LIT5 Martiny et al., 2005 Martiny N., R. Santer, I. Smolskaia,
Vicarious calibration of MERIS over dark
waters in the near infrared, Remote Sensing
of Environment 94 (2005) 475–490
LIT6 Vermote et al., 2008 6SV manual,
http://6s.ltdri.org/6S_code2_thiner_stuff/6s_l
tdri_org_manual.htm
LIT7 Thuillier et al., 2003 Thuillier, G., M. Hersé, P. C. Simon, D.
Labs, H. Mandel, D. Gillotay, and T. Foujols,
2003, "The solar spectral irradiance from 200
to 2400 nm as measured by the SOLSPEC
spectrometer from the ATLAS 1-2-3 and
EURECA missions, Solar Physics, 214(1): 1-
22
LIT8 Gordon and Wang, 1994 Howard R. Gordon and Menghua Wang,
"Retrieval of water-leaving radiance and
aerosol optical thickness over the oceans with
SeaWiFS: a preliminary algorithm," Appl.
Opt. 33, 443-452 (1994)
LIT9 Shettle and Fenn, 1979 Shettle, E. and R.W. Fenn, Models for the
Aerosols of the Lower Atmosphere and the
Effects of Humidity Variations on their
Optical Properties, AFGL-TR-79-0214,
Environmental Research Paper, No. 676
LIT10 Morel, 1988 A. Morel, Optical modeling of the upper
ocean in relation to its biogenous matter
content (case I
waters), Journal of Geophysical Research,
93(C9), 10749-10768, 1988
LIT11 Fougnie et al., 2002 Fougnie, B., P. Henry, A. Morel, D. Antoine
and F. Montagner, Identification and
Characterization of Stable Homogeneous
Oceanic Zones: Climatology and Impact on
In-flight Calibration of Space Sensor over
Rayleigh Scattering, Proceedings of Ocean
Optics XVI, Santa Fe, New Mexico, 18-22
November 2002
LIT12 Kotchenova and Vermote, 2007 Svetlana Y. Kotchenova and Eric F.
Vermote, "Validation of a vector version of
the 6S radiative transfer code for atmospheric
correction of satellite data. Part II.
Homogeneous Lambertian and anisotropic
surfaces," Appl. Opt. 46, 4455-4464 (2007)
LIT13 Gordon et al., 1988 Gordon, H. R., O. B. Brown, R. H. Evans, J.
W. Brown, R. C. Smith, K. S. Baker, and D.
K. Clark (1988), A semianalytic radiance
model of ocean
color, J. Geophys. Res., 93, 10,909– 10,924.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 13/139
LIT14 Wang, 2005 M. Wang, A refinement for the Rayleigh
radiance computation with variation of the
atmospheric pressure, Int. J. Remote Sens.
26, 5651–5653 (2005).
LIT15 Tanré et al., 1990 D. Tanré, C. Deroo, P. Duhaut, M. Herman,
J. J. Morcette, J. Perbos and P. Y.Deschamps,
Description of a computer code to simulate
the satellite signal in the solar spectrum: 5S
code. Int. J. Remote Sensing,11,659-668
(1990).
LIT16 Rahman and Dedieu, 1994 Rahman H., Dedieu G. 1994: SMAC : a
simplified method for the atmospheric
correction of satellite measurements in the
solar spectrum Int. J. Remote Sensing,
vol.15, no.1, 123-143.
LIT17 Teillet et al, 2007 Teillet, P.M., J.A. Barsi, G. Chander, and
K.J. Thome (2007). Prime Candidate Earth
Targets for the Post-Launch Radiometric
Calibration of Space-Based Optical Imaging
Instruments. Proc. SPIE 6677, San Diego,
2007
LIT18 Biggar et al., 2003 Stuart F. Biggar, Kurtis J. Thome, and Wit
Wisniewski,
Vicarious Radiometric Calibration of EO-1
Sensorsby Reference to High-Reflectance
Ground Targets, IEEE TRANSACTIONS
ON GEOSCIENCE AND REMOTE
SENSING, VOL. 41, NO. 6, JUNE 2003.
LIT19 Thome, 2001 K.J. Thome, 2001, Absolute radiometric
calibration of Landsat 7 ETM+ using the
reflectance-based method, Remote Sensing
of Environment 78 (2001) 27– 38
LIT20 Dinguirard and Slater, 1999 Calibration of Space-Multispectral Imaging
Sensors: A Review, Remote Sens. Environ.
68:194-205 (1999).
LIT21 Biggar et al., 1994 Biggar S.F., Slater P.N., and Gellman D.I.,
Uncertainties in the in-flight calibration of
sensors with reference to measured ground
sites in the 0.4 to 1.1 µm range. Rem. Sens.
Environ. 48: 245-252
LIT22 Itten et al., 2008 Itten, K.I.,. F. Dell’Endice, A. Hueni , M.
Kneubühler,D. Schläpfer, D. Odermatt, F.
Seidel, S. Huber, J. Schopfer, T.
Kellenberger, Y. Bühler, P. D’Odorico, J.
Nieke, E. Alberti, and Koen Meuleman,
APEX – the Hyperspectral ESA Airborne
Prism Experiment, Sensors 2008, 8, 6235-
6259
LIT23 Biesemans et al., 2007 Biesemans J., Sterckx, S., E. Knaeps, K.
Vreys, S. Adriaensen, J. Hooyberghs, K.
Meuleman, P. Kempeneers, B. Deronde, J.
Everaerts, J. Schlaepfer and J. Nieke (2007) ;
Image Processing Workflows For Airborne
Remote Sensing. Proceedings 5th EARSeL
Workshop on Imaging Spectroscopy. Bruges,
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 14/139
Belgium ; eds. I. Reusen.
LIT24 Nieke, 2001 Nieke J., Technical Note 1 Compilation of
Scientific Variables, Specification of an
airborne spectroradiometer to calibrate and
validate spaceborne Earth Observation
sensors, 2001
LIT25 Teillet et al., 2001 P.M. Teillet, G. Fedosejevs, R.P. Gauthier,
N.T. O’Neill, K.J. Thome, S.F. Biggar, H.
Ripley, A. Meygret, A generalized approach
to the vicarious calibration of multiple Earth
Observation sensors using hyperspectral data,
Remote Sensing of Environment 77 (2001)
304-327
LIT26 Hovis et al., 1985 W.A. Hovis, J.S. Knoll, G.R. Smith; Aircraft
measurements for calibration of an orbiting
spacecraft sensor; Appl. Optics; Vol. 24; pp
407-410 (1985)
LIT27 Hagolle et al., 2004 Hagolle, O., Nicolas, J.-M., Fougnie, B.,
Cabot, F. and Henry P. (2004)..Absolute
calibration of VEGETATION derived from
an interband method based on the Sun glint
over ocean. IEEE Transactions on
Geoscience and Remote Sensing, 42(7): 1472
– 1481.
LIT28 Cox and Munk,1954 Cox and Munk, Measurement of the
roughness
of the sea surface from photographs of the
Sun’s
glitter, J. Opt. Soc. Am., 44, 838–850, 1954
LIT29 Born and Wolf,1975 Born, M. and Wolf, E., 1975. Principles of
Optics. In: , Pergamon, Oxford, p. 90.
LIT30 Hale and Querry,1973 Hale and Querry,1973. Optical Constants of
Water in the 200-nm to 200-µm Wavelength
Region," Appl. Opt. 12, 555-563,1973.
LIT31 Lafrance et al., 2002 Lafrance B., Hagolle O., Bonnel B., et al.,
2002 :
Interband calibration over clouds for
POLDER space sensor. IEEE Trans Geosc.
Rem. Sens., 40 (1): 131-142
LIT32 Henry and Meygret, 2001 Henry and Meygret 2001. Calibration of
HRVIR and VEGETATION cameras on
SPOT4. COSPAR 2000. Advances in Space
Research 28 1 (2001), pp. 49–58.
LIT33 Sohn et al., 2009 Sohn B, Ham S,Yang P (2009) Possibility of
the Visible-Channel Calibration Using Deep
Convective Clouds Overshooting the TTL.
Journal of Applied Meteorology and
Climatology 48(11): 2271
LIT34 Hu and Stamnes,1993 Hu, Y.X. and K. Stamnes, An Accurate
parameterization of the Radiative Properties
of Water Clouds Suitable for Use in Climate
Models. J. Climate, 6, 728--742, 1993.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 15/139
LIT35 Baum et al,2005 Baum, B. A., A. J. Heymsfield, P. Yang, and
S. T. Bedka, ’Bulk scattering models for the
remote sensing of ice clouds. Part 1:
Microphysical data and models’, 2005,
J.Appl. Meteor., 44, 1885-1895
Baum, B. A., P. Yang, A. J. Heymsfield, S.
Platnick, M. D. King, Y.-X. Hu, and S. T.
Bedka, ’Bulk scattering models for the
remote sensing of ice clouds. Part 2:
Narrowband models’, 2005, J. Appl. Meteor.,
44, 1896-1911.
Baum, B. A., P. Yang, S. L. Nasiri, A. K.
Heidinger, A. J. Heymsfield, and J. Li, ’Bulk
scattering properties for the remote sensing
of ice clouds. Part 3: High resolution spectral
models from 100 to 3250 cm-1.’, 2007, J.
Appl. Meteor. Clim., Vol. 46,423-434.
LIT36 Zhang et al., 2009 Zhang, Z., P. Yang, G. W. Kattawar, J. Riedi,
L. C-Labonnote, B. A. Baum, S. Platnick,
and H. L. Huang, 2009: Influence of ice
particle model on satellite ice cloud retrieval:
lessons learned from MODIS and POLDER
cloud product comparison. Atmos. Chem.
Phys., 9, 7115-7129
LIT37 Lattanzio et al., 2006 Lattanzio, A., P. D. Watts, and Y. Govaerts,
2006: Activity report on physical
interpretation of
warmwater vapour pixels, EUMETSAT,
Technical Memorandum 14.
LIT38 Cosnefroy et al., 1996 Cosnefroy, H., M. Leroy, and X. Briottet,
1996:
Selection and characterization of Saharan and
Arabian desert sites for the calibration of
optical
satellite sensors, Reomte Sensing Env., 58,
101-
114
LIT39 Rahman et al., 1993 Rahman, H., Pinty, B. and Verstraete, M. M.
(1993) Coupled surface-atmosphere
reflectance (CSAR) model 2. Semiempirical
surface model usable with NOAA Advanced
Very High 93 Resolution Radiometer data.
Journal of Geophysical Research, 98, pp.
20791-20801.
LIT40 Stone, 2008 Stone, 2008. Use of the Moon for in-flight
calibration stability monitoring QA4EO-
WGCV-IVO-CLP-001 (CEOS Cal/val
portal) (2008).
LIT41 Govaerts et al., 2004 Govaerts, Y. M., M. Clerici, et al. (2004).
"Operational Calibration of the Meteosat
Radiometer VIS Band." IEEE Transactions
on Geoscience and Remote Sensing 42(9):
1900-1914.
LIT42 Bruegge et al., 1999 C. Bruegge , N. Chrien, D. Diner ATBD-
MISR-02 MISR: Level 1 In-Flight
Radiometric Calibration and Characterization
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 16/139
December 13, 1999
LIT43 Levenson et al, 2000 Levenson, M.S. et al: An approach to
combining results from multiple methods
motivated by the ISO GUM, J. Res Natl.
Stand. Technol. 105, 571 (2000)
LIT44 Galbraith et al. Amy Galbraith, James Theiler, and Steve
Bender. Resampling Methods for the MTI
Coregistration Product. Proceedings SPIE
5093.
LIT45 Press et al, 2007 Press W.H., Teukolsky S. A., Vetterling
W.T., Flannery B.P., 2007: Numerical
recipes: the art of scientific computing
LIT46 Fougnie et al, 2000 Fougnie, B., Henry, P., Cabot, F., Meygret,
A. and Laubies, M.-C, 200, Vegetation: in-
flight multi-angular calibration, in Earth
Observing Systems V, Proc. SPIE 4135, 331-
338
LIT47 Fox, 2004 Fox, N. Validated data and removal of bias
through traceability to SI units, Post-Launch
Calibratin of Satellite Sensors – Morrain &
Budge (eds), 2004 Taylor & Francis group
LIT48 Zhang et al., 2000 Zuxun Zhang, Jlanqlng Zhang, Mlngsheng
Uao, and U Zhang. Automatic Registration of
Multi-Source Imagery Based on Global
Image Matching. Photogrammetric
Engineering & Remote Sensing
Vol. 66, No. 5, May 2000, pp. 625-629.
LIT49 Canters et al., 2009 J. Ma , J.C.-W. Chan , F. Canters. Automatic
image registration of multi-angle imagery for
CHRIS/PROBA. 6th EARSeL SIG IS
workshop, March 16-19, 2009, Tel
Aviv, Israel
LIT50 Tempelmann et al. Udo Tempelmann, Ludger Hinsken, Utz
Recke. ADS40 Calibration & Verification
Process. Proceedings of the ISPRS WG1/5
International workshop.
LIT51 Hagolle and Cabot, 2002 Hagolle O;, Cabot F.: Absolute calibration of
MERIS using natural targets. Proceedings of
“MERIS level 2 validation workshop,
Frascati, Italy, DeC. 13-14, 2002.
LIT52 Hagolle and Cabot, 2003 Hagolle O. , Cabot F., Absolute Calibration
Of Meris Using Natural Targets, MAVT
2003 proceeding.
LIT53 Morel and Gentili, 2009 Morel A, Gentili B.,A simple band ratio
technique to quantify the colored dissolved
and detrital organic material from ocean
color remotely sensed data, Remote Sensing
of Environment,2009, 113:998-1011
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 17/139
LIT54 Morel et al. 2010 Morel A., Claustre H. and Gentili, B. . The
most oligotrophic subtropical zones of the
global ocean: similarities and differences in
terms of chlorophyll and yellow substance.
Biogeosciences Discuss. ,2010,7: 5047-5079.
LIT55 Morel et al.,2007 Morel A., Claustre H., Antoine, D. abd
Gentili, B. Natural variability of bio-optical
properties in Case 1 waters: attenuation and
reflectance within the visible and near-UV
spectral domains as observed in South Pacific
and Mediterranean waters.
Biogeosciences,2007,4:913-925.
LIT56 Fougnie et al.,2010 Bertrand Fougnie, Jérome Llido, Lydwine
Gross-Colzy, Patrice Henry and Denis
Blumstein, "Climatology of oceanic zones
suitable for in-flight calibration of space
sensors", Proc. SPIE 7807, 78070S (2010);
doi:10.1117/12.859828
LIT57 Kieffer et al., 2002 H.H. Kieffer, P. Jarecke, and J. Pearlman,
"Initial Lunar Calibration Observations by
the EO-1 Hyperion Imaging Spectrometer",
Proc. SPIE 4480, 247-258 (2002);
LIT58 Stone et al., 2013 Stone, T.C., W.B. Rossow, J. Ferrier, and
L.M. Finkelman, 2013: Evaluation of ISCCP
multisatellite radiance calibration for
geostationary imager visible channels using
the Moon. IEEE Trans. Geosci. Remote
Sens., 51, 1255-1266
Table 2: Literature references
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 18/139
3. TERMS, DEFINITIONS AND ABBREVIATED TERMS
General terms, definitions and abbreviations used within the scope of this document are listed in
[PVDOC-303] Directory of Acronyms and Abbreviations.
Abbreviations and acronyms used within the scope of this document are listed in Table 3:
6SV Second Simulation of a Satellite Signal in the Solar Spectrum,
Vector
AGL Above ground level
AOT Aerosol Optical Thickness
CCD Charge Coupled Device
CMOS Complementary Metal Oxide Semiconductor
DCC Deep Convective Clouds
FWHM Full Width at Half Maximum
IVOS Infrared and Visible Optical Systems
LST Local Solar Time
PDF probability distribution function
ROLO Robotic Lunar Observatory
RVP Rahman-Pinty-Verstraete
WGCV Working Group Calibration/Validation
Table 3: Abbreviations and acronyms
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 19/139
4. RADIOMETRIC CALIBRATION
4.1 Introduction
The assessment of the PROBA-V performance, the analysis of the image quality and the
calibration after launch will be performed by the PROBA-V Image Quality Center (IQC). Due to
outgassing phenomena during launch, aging of the optical parts and cosmic ray damage variation
in the characteristics of PROBA-V instrument are likely to occur in orbit, necessitating the need
for in-orbit stability monitoring and calibration. The Image Quality Center will ensure the highest
possible image quality, both radiometrically and geometrically. Given the constraints on power
consumption and the small size and weight of the platform (160 kg) vicarious calibration
techniques will be used to monitor sensor performance over time.
In this document the algorithms used by IQC to monitor the sensor performance and to determine
the parameters to be supplied to the processing facility (PF) are described in detail. Related
documents are : [PVDOC-624] which gives the details of the processing to be applied to the
PROBA-V images ; The calibration plan [PVDOC-615] which contains details on calibration
zones, frequency of calibration and instrument setting .
To cover the wide angular field of view (101°) the optical design of PROBA-V is made up of
three cameras (identical Mirror Anastigmatic (TMA) Telescopes). Each camera has 2 focal
planes, one for the short wave infrared (SWIR) band and one for the visible and near-infrared
(VNIR) bands. The VNIR detector consists of 4 lines of 5200 pixels. Three spectral bands are
selected, compatible with SPOT-Vegetation: BLUE, RED and NIR. The SWIR detector is a
linear array composed of 3 mechanically butted detectors of 1024 pixels. The correspondence
between instrument design and captured image data is schematically represented in Figure 1.
instrument = 3 camera’s
camera = 2 sensors
VNIR sensor =3 detectors (blue,red,nir)
=3 strips
SWIR sensor =1 detector (swir)
=3 butted strips
scene
image = N bands
(detector|image) pixel
(i,j) = (column, row)
frame = M rows
(M=1 for a line sensor)
j
i
image pixel = location (i,j[,band])
Figure 1: Conceptual overview
The three PROBA-V cameras are treated separately in the in-flight radiometric calibration, except
in section (4.7) where possible biases between cameras are analysed.
In section 4.2 the radiometric model is given. It describes the conversion of the digital output of a
pixel to the at-sensor spectral radiance. In this section special attention is given to temperature
effects due to lack of an active thermal control mechanism. The conversion of digital numbers to
spectral radiances is achieved through the use of the different sensor calibration coefficients.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 20/139
These are: dark currents, absolute calibration coefficients and equalization/non-uniformity/multi-
angular coefficients. The initial values of the calibration parameters is fixed by the pre-launch
calibration measurements. After launch these parameters are monitored, validated and, if needed,
updated by the vicarious calibration activities described in the subsequent sections.
In section 4.3 first the algorithms to define the dark currents per pixel are described (section
4.3.1). Degraded or defect pixels can be detected using the algorithms given in section 4.3.2. As a
linear radiometric model is assumed, it is essential that any deviation from this assumed linearity
can be detected. The in-flight approach for detecting deviations from non-linearity is described in
4.3.3.
In the next three section (4.4, 4.5 and 0) the methods for in-flight calibration of the absolute
calibration coefficients are given. Several different and independent approaches will be used as
The methods are not always suitable for all bands and combination of methods are
needed to allow accurate calibration of all spectral bands
Independent validation of the results is required to determine and to account for
systematic errors in one or more techniques. For instance uncertainties in the
characterization of target BRDF or assumptions in aerosol characteristics can induce
systematic errors (Govaerts and Clerici., 2004 [LIT1]).
Some methods can only be used during a few months in the year (eg. calibration over
Antarctica)
One method is not sufficient to cover the full dynamic range of the sensor : bright (e.g.
deserts) versus dark calibration areas (e.g. oceans)
The uncertainty in the calibration results can be decreased by consistency check of the
different methods the procedures are often limited to a few spectral bands
A distinction has been made between absolute (4.4) and relative calibration methods (4.5). The
operational absolute calibration for BLUE and RED PROBA-V bands can be performed using the
so-called Rayleigh calibration approach (4.4.1). Other approaches are the reflectance based
methods (4.4.2) and underflights with the APEX sensor (4.4.3). Due to the high operational costs
for these latter two methods they can’t however be used frequently, and therefore they are mainly
considered for validation purposes. The results of the Rayleigh calibration method can be
transferred to other bands (NIR, SWIR) based on a ‘relative’ inter-band calibration approach that
uses almost spectrally invariant targets as sun glint (4.5.1) or deep convective clouds (which is
not suitable for SWIR) (4.5.2). Temporally stable targets are ideal for monitoring the stability of
PROBA-V. Suitable targets for multi-temporal calibration are stable deserts (4.5.3), Antarctica
(4.5.4) and even the moon (4.5.5) .
To reduce both random and systematic error effects, calibration coefficients derived over a large
number of images and obtained with different methods will be statistically averaged. The
statistical approach used to determine the absolute calibration coefficient to be used by the
processing facility is detailed in 0.
In the absolute and relative calibration methods the three cameras are treated separately, which
may introduce biases between the cameras. In the overlap zone, targets are simultaneously seen
by 2 independent cameras. In 4.7 it is described how this overlap zone can be used to indicate
possible bias between cameras.
The in-flight monitoring and calibration of the equalization coefficients is described in 4.8. The
equalization coefficient can be split up in a low and a high frequency term. The in-flight
determination of the low frequency is performed using stable deserts with know bidirectional
effects. High frequency variation can be assessed using images over Antarctica or Greenland
(4.8.3.4).
The algorithms for assessment of the general instrument image quality are given in section 4.9.
The section describes noise analyses (4.9.1) , MTF measurement (4.9.2) and the quantification of
SNR (4.9.3).
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 21/139
In section 4.10 is described how these algorithms will be validated before and after launch. In
order to secure proper data continuity and consistency between SPOT-VGT and PROBA-V cross
sensor calibration is essential. The necessity of almost simultaneous observations can be
overcome by the use of stable sites as deserts or Antarctica. This approach is described in 4.10.1.
This cross-sensor calibration is also considered as independent validation of routine calibration.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 22/139
4.2 Sensor Radiometric Model
4.2.1 Model overview
The Sensor Radiometric Model defines the relation between the raw digital output which is
registered by the sensor and sent down for data processing, and the derived effective spectral
radiance assumed to be present at the sensor. This model will be used in two stages:
pre-launch calibration: at this stage, both the digital output and the input effective
spectral radiance are known up to a given accuracy, and the calibration parameters
determining the relation between both can be derived;
post-launch calibration: at this stage, the effective radiance must be derived from the
digital output and the calibration parameters currently used. The initial values of the
calibration parameters is fixed by the pre-launch calibration measurements, these
parameters are then updated and validated by the vicarious calibration activities
described in the subsequent sections.
The mathematical model representing the radiometric response of PROBA-V is :
k
g
kk
gi
k
T
kk
gi
kk
Tgi
k
g
k
Tgi NLdITITgALoffdITITdcAVGDN ,,,,,, .
with
k denotes the spectral band;
i is the pixel index;
g is the gain setting;
T is the detector temperature;
DN is the digital number.
where
kTgidc ,, [units : LSB/s] is the dark current coefficient. This coefficient depends on the
temperature and for the SWIR detectors it is also gain setting dependent
IT the integration time
kdIT [units : s] is the so-called integration time offset, a parameter introduced allowing
to improve the radiometric model fit with calibration results. This parameter is
dependent on the spectral band k.
kgioff , [units: LSB] is the offset which is both dependant on and varying in proportion
with the average gain level
kTA [units: LSB W
-1 m² µm Sr s
-1]: is the absolute radiometric coefficient, indicating the
sensitivity of the spectral band k. It is dependent on the detector temperature.
k
gAVG [unitless] is the average gain level (average over all recorded pixels i). By
definition, is set to 1 for the nominal gain setting, and best fitted by calibration for the
other gain settings.
kgNL [units: LSB] is a non-linear function of DN, expressed in DN units (i.e. LSB
counts). It can be expressed as a look-up table for each pixel i of each spectral band k
and for gain g.
kgig , [unitless] is the equalisation coefficient and represents the pixel to pixel response
variation due to the possible pixel to pixel gain variation, so kgg , being the average of
kgig , over all recorded pixels i, is by definition 1.
4.2.1.1 Temperature effect on dark current
The dark current is higly dependent on temperature following an exponentional model :
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 23/139
referenceb
k
ireference
k
gi
k
giTTk
ETdcTdc
11exp).()( ,,
Equation 1: Analytic relation dark current-temperature
with kiE : the activation energy expressed in J units
bk : Boltzmann constant
and Treference : the reference Tempature (expressed in Kelvin), arbritary chosen to be 263.15 K
(i.e. -10°C).
Dark current calibration will be performed in-flight during night at a certain temperature, T =
Tactual. Following the algorithms given in 4.3.1.2 and Equation 1 )(, reference
k
gi Tdc is calculated in
the IQC on the basis of these measurements.
In the processing performed by PF Equation 1 is not explicitly applied, but instead the IQC
generates a LUT on the basis of Equation 1 with for a number of Tempatures (T) the
corresponding )(, Tdck
gi . This LUT is stored in the ICP files. By linear interpolation the values at
surrounding temperatures can be generated.
On the basis of the available pre-flight/onground calibration data the LUT will contain the
)(, Tdck
gi for the following set of Temperature values : 273, 268, 263, 258 and 253 K.
Based on the pre-flight ICP data the error made by linear interpolation is negligible as illustrated
below.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 24/139
From the current thermal design analysis, it appears that the maximum change in temperature
expected over one orbit is below 2 °C for the VNIR band, below 5°C for the SWIR band.
Differences caused by temperature is taken into account through the implementation of this LUT
In addition, over the course of a year there is a seasonal effect that results in a temperature
difference. Based on the Hot/Cold case analysis done by QS, a worst-case monthly variation of 1
to 1.5°C is expected.
4.2.2 Inverse model
In calibration operations, the model has to be used in inverse direction, starting from DN data and
working toward effective spectral radiance. In radiometric characterization approaches, three
phases can be distinguished.
As a first phase, the acquired DN result must be corrected for the average gain level (for nominal
gain this is 1), nonlinearity and for the offset:
k
gAVG))/(DNNL- (DN k
acquiredi,
k
g
k
acquiredi, kgioff ,
The second phase is a correction for dark current (section 4.3.1). The objective here is to deduce
and then subtract the dark signal offset -taking into account the integration time and integration
time offset- and retrieve an intermediate DN result:
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 25/139
kk
Tgi
k
i
k
igcor dITITdcDNDN ,,,1
Equation 2: Dark current offset correction
In the third phase the current values for <A,g,.> are used to derive the effective spectral radiance
result:
kk
gi
k
T
k
igcork
iTOAdITITgA
DNL
,
,1
,
Equation 3: Conversion to effective spectral radiance
These two equations can be put together to describe the relation between acquired DN data and
effective spectral radiance:
kk
gi
k
T
kk
Tgi
k
gk
iTOAdITITgA
dITITdcAVGL
,
,,
k
gi,
k
acquiredi,
k
g
k
acquiredi,
,
off-))/(DNNL -(DN
Equation 4: Inverse sensor model relation
4.2.3 Operational dependencies
Each camera will be treated separately. Estimation of the different parameters in the sensor model
is discussed in subsequent sections:
dark current kidc in section 4.3.1.
absolute calibration coefficient kA in sections 4.4, 4.5 , 0
equalization coefficient kig in section 4.8
nonlinearity parameter NL is not estimated in-flight, but the linearity of the system after
nonlinearity correction is verified by the process described in section 4.3.3.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 26/139
4.3 Radiometric Characterization
4.3.1 Dark Current
4.3.1.1 Introduction
Dark current (dci,gk) is caused by thermally generated electrons that build up in the pixels. The
magnitude of the dark current is expected to increase with time due to space radiation. Moreover,
as described in section 4.2.1.1, noticeable variations of dark current are expected over the course
of the year as a result of temperature effects. It is therefore important to monitor the dark current
in orbit.
Uncompressed images taken during the nighttime portion of the orbit of dark ocean sites (situated
in Northern hemisphere in winter and Southern hemisphere during summer) will be used to
determine the dark current values for all pixels.
Central assumption in the calibration is that the effective signal captured during the dark current
calibration is much lower than the dark current itself. The effective signal of each pixel will then
be composed mainly of its dark current signal, the associated dark current shot noise and the
signal-independent read noise. The latter two contributions can be eliminated by taking the mean
of the signal of each pixel over time.
4.3.1.2 Algorithm
Starting from the assumptions of the previous section, the steps to follow in this algorithm are:
Select dark ocean sites (at night time) for which the effective signal can be expected to
be sufficiently low. Sites have been selected and defined in the In-flight Characterization
and Calibration Plan [PVDOC-615]. For these sites, L1B data is retrieved for regular
overpasses of at least 500 lines.
Remove saturated values1 from further processing.
Correct the acquired DN ( kacquiredi,DN ) for the average gain level (for nominal gain this is
1),Non-linearity and offset : k
gi,
k
acquiredi,
k
g
k
acquiredi,
k
i off-))/(DNNL -DN(DN k
gAVG
Next, the DN values are and normalized by the Integration time:
k
actual
k
gi dITITTdc /DN)( k
i,
)(, reference
k
gi Tdc is calculated as :
referenceactualb
k
iactual
k
gireference
k
giTTk
ETdcTdc
11exp/)()( ,,
with actualT the actual temperature of the dark current measurements performed over the
oceans.
For each pixel, lines containing significant outliers are detected. Significant line outliers
are an indication of ‘local events’, either caused by an unwanted signal (such as moon
glint) or unwanted behaviour of the instrument. Line outliers can be detected and
removed by determining the median and median absolute deviation. This approach is
discussed in 4.3.1.2.1.
When line outliers are removed, the average value is calculated over the remaining lines l
lreference
k
lgireference
k
gi TdcTdc )()( ,,, . The result is considered as the current value of dark
current/s for this pixel and gain setting.
1A DN value is saturated when the value is equal to the maximum bit of the data.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 27/139
The time range corresponding to 500 lines is 7,5 s.
For the optical band signals, all dark current values, not including saturated lines, can be
considered reliable. For the SWIR signals however, a stronger influence of dark current and risk
of radiative damage must be accounted for. This is done by performing a quality check procedure
on the averaged pixel values (see 4.3.1.2.2). The result of this procedure is a set of dark current
values for the qualified pixels, and a quality map allowing for four cases:
‘undefined’: the pixel is saturated, and no value could be defined for it
‘aberrant’: value of the pixel is suspect, and no reliable value could be defined for it
‘singular’: value of the pixel is suspect, but a corrected value is defined by looking at
neighbouring values
‘good’: the pixel’s value is not suspect.
4.3.1.2.1 Line outlier detection
The concepts of median and median absolute deviation (MAD) are better suited to the detection
of outliers than the mean value and standard deviation. The calculation is defined as:
lkgi
klgi
kgi
lk
lgikgi
MEDDNMAD
DNMED
)median(
)median(
,,,,
,,,
Equation 5: Median metrics for outlier detection
The MAD value can be made equivalent to a standard deviation of a distribution by multiplying
with a constant K. In the assumption of a normal distribution, K=1.4826. Hence: kgi
kgiMAD
MADS ,,,.48261
Equation 6: Equivalent median-standard deviation for normal distributions
When the MAD value is zero, most values are close to the median. In this case no outliers are
eliminated, since no additional information on the values differing from the median has been
retrieved by the statistics.
For every non-zero value, the 2-sigma confidence interval is selected as the region of valid lines.
Every other line is considered an outlier. The DN values therefore have to obey the following
condition: k
giMADkgi
klgi
kgiMAD
kgi SMEDDNSMED
,,,,,,,, .. 961961
Equation 7: 2-sigma confidence interval based on median-standard deviation
4.3.1.2.2 SWIR pixel quality check and pixel outlier correction
Pixel quality is checked and if possible and necessary corrected by comparing the difference in
value between a pixel and its neighbours. To avoid the influence of saturated pixels in this
calculation, a pixel mask sat(j) is defined such that sat(j) = 1 when the pixel is saturated in all
lines, and sat(j) = 0 when the pixel is not saturated.
A status list is defined, containing a status value for all pixels. Possible status values are:
blank: no value defined yet
undefined, aberrant, singular, good: as described above.
All values from first pixel 1 to final pixel Np are set to blank initially. At either edge, an extra
value is given and set to aberrant to account for the fact that the edge pixels have only one
neighbour they can compare to. In addition, the status of saturated pixels is set to undefined.
These operations are illustrated in Figure 2.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 28/139
Figure 2: Initial quality status in red for 30 pixels.
Triangles at high levels represent ‘undefined’ status, triangles at low levels represent ‘blank’
status. The red crosses at either end represent ‘aberrant’ status for the extra edge values.
The quality status of pixels is derived from a statistical analysis of their averaged value (which is
the metric of the dark current value for ‘good’ pixels). This can be motivated by observing Figure
3.
Figure 3: Dark current data for VGT-MIR data (from 512 lines)
This shows dark current data extracting from the CNES SPOT-VGT mission, for 512 lines of the
MIR band (which corresponds spectrally to PROBA-V SWIR). The values are sorted from small
to large. This shows that the large majority of pixels displays a very similar dark response, with a
minor set of outliers found at either end of the signal range. A pixel’s status can therefore be
defined as ‘good’, ‘singular’ or ‘aberrant’ depending on its deviation from the mean value.
An ‘aberrant’ status is mainly seen at the higher signal end: these are associated with pixels
showing a significantly different behaviour, and are therefore assumed to be impossible to
correct. A ‘good’ status is associated with pixels who are very close to the mean. ‘Singular’ status
is to describe intermediaries, which might be corrected by comparing with neighbouring values.
Two metrics are calculated as demarcations for these three categories. The MAD-standard
deviation is again calculated, as a conservative measure to distinguish ‘good’ pixels from
‘singular’ ones. A less tight measure is the standard deviation: this is used to distinguish
‘singular’ pixels from ‘aberrant’ pixels. The demarcations are described as such:
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 29/139
ik
gik
gi dcmeandc )( ,, Status
SWIRgMAD,
.... 961 good
SWIRgstd
SWIRgMAD ,,
..... 961961 singular
SWIRgstd ,
.... 961 aberrant
Table 4: Pixel quality calculation
Once the quality of all pixels is assessed, a correction can be performed for every singular pixel.
This is done by checking the status of its neighbours. If one of its neighbours is ‘good’, it is
assumed that the same dark current is present in both pixels. If not, then the ‘singular’ value is not
corrected.
4.3.2 Detection of defective and degraded pixels
4.3.2.1 Introduction
Defective pixels can either be stuck pixels (which exhibit consistently a lower or higher bias of
charge) or they can have a gain reduction, deviating to such an extent that a calibration by
equalization is no longer possible for these pixels. Defects can develop from radiation damage in-
flight or unexpected malfunctioning, and therefore the status of every pixel needs to be assessed
continuously in-flight. Aside from a supervised evaluation on uniform sites done at the IQC at a
regular basis (the regularity of which is to be decided during the commissioning phase), an
automatic method of correction is described but not implemented in the PF [PVDOC-624]. This
method allows the PF to catch new defects in a fast way, without having to rely on an update of
the pixel statuses by the IQC. This type of algorithm will only be implemented in the PF if the
amount and frequency of detected defects is too much to rely on supervised evaluation alone.
During the commissioning phase, the supervised evaluation of pixel quality is intensely
investigated to gain confidence that the followed procedure maintains control and overview of
quality status of every pixel. The investigation proceeds according to several phases which are
described in sections 4.3.2.2.3-4.3.2.2.6. When the investigation is complete, the nominal
supervised evaluation is set as fixed for the nominal operations.
4.3.2.2 IQC evaluation activities
The evaluation focuses on the imaging of uniform sites, as is done in other calibration activities.
Strong deviations from a uniform signal are then expected to be caused by the pixel’s behaviour
only, therefore allowing aberrant behaviour to be detected more easily.
Because of radiometric noise effects, a good grasp of a pixel’s systematic behaviour can only be
done when averaging out the pixel’s signal over a sufficient amount of lines. However, one must
be cautious for unexpected ‘local’ events unrelated to pixel defects. These events can be avoided
by detecting outliers. This process is described in 4.3.2.2.1.
During commissioning, the development of the defective pixels evaluation proceeds according to
several phases. In the first phase, detections are performed for each uniform site separately. This
allows to detect bias deviations specific for a site (deviations of averaged pixel signals from the
averaged signal of that uniform site). The detection results can be compared across different sites
and sites can thus be validated. This process is described in section 4.3.2.2.3.
After validation, different sites are integrated into one detection system. The goal is here is to
combine information from different sites, and deliver one output. This is described in section
4.3.2.2.4.
At this point, several pixels can still be flagged as suspect. This might be because of defects not
explainable as a strong bias effect. A gain method is then used as an additional detection strategy.
This is discussed in section 4.3.2.2.5.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 30/139
Finally, the final combination of all relevant sites and methods is integrated, and fixed as the final
IQC evaluation process. As data of relevant sites are also used by other calibration methods, these
relations are briefly discussed. This is done in section 4.3.2.2.6.
4.3.2.2.1 Line outlier detection & averaging
The goal of line outlier detection is to detect temporary deviations of a pixel’s signal that are not
typical for that pixel. Causes for this are unexpected events, such as clouds, a local reflection (sun
glint, moon glint), or radiative collision. Radiative collision are possible causes for permanent
defects, but can also temporarily influence the pixel’s signal; only the latter is considered a line
outlier.
Obvious line outliers are saturated values. These are removed from further processing. Note that
by removing saturated values, it could be possible that the remaining lines are too few to be
reliable for averaging. A number of unsaturated lines (default = 200) is required at minimum, if
this is not the case, the pixel status calculation is considered as failed and the pixel is flagged as
‘suspect’(see 4.3.2.2.2). If all lines are consistently saturated , for a site containing at least a
minimum number of lines for this pixel (default = 200), this is considered a saturated line and the
pixel is flagged as ‘bad’.
Next, the acquired DN ( kacquiredi,DN ) for Non-linearity and offset :
k
gi,
k
acquiredi,
k
g
k
acquiredi,
k
i off-))/(DNNL DN(DN k
gAVG
Further line outliers are detected as deviations from the mean signal. This approach is similar to
that in section 4.3.1.2.1 and Equation 7:
k
giMADkgi
klgi
kgiMAD
kig SMEDDNSMED
,,,,,,,.. 961961
In this case however, it is not optimal to use the concept of median deviation(MAD), because a
defective pixel can have a lot of erratic lines, but still have a good mean and median2. Therefore,
the MAD is replaced by a standard deviation figure, and the median by the mean figure:
k
giSTDkgi
klgi
kgiSTD
kgi DN
,,,,,,,, .. 961961
When kgiSTD ,, is significantly larger than the average noise value for other pixels, the pixel
status calculation is considered as failed and the pixel is flagged as ‘bad’ (see 4.3.2.2.2).
The averaged pixel value is then calculated by averaging over the remaining lines:
l
klgi
kmeangi DNDN
,,,,
4.3.2.2.2 Pixel status definitions
Three statuses are distinguished:
‘good’ : pixels are qualified as good when their behaviour is identified as normal
‘bad’ : pixels are qualified as aberrant when their behaviour is defective, not normal
‘suspect’ : pixels are qualified as suspect when their behaviour has been impossible to
identify. When a quality detection has failed or was impossible to determine, a pixel is
qualified as suspect.
2 Example: Take as reference of a good pixel, one with values averaged around 1000DN, and some small
noise (values of 1005, 994, 1002,..). A bad pixel is then the following: an average of 1000DN, with 200
lines at value 100 and 800lines at 1225. This behavior is bad, considering that the imaged site is
supposedly uniform, and should not return values of 100 at one time, 800 at another. Still, the MAD
figure will be small, because the erratic behavior of the 200 lines will not be detected.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 31/139
4.3.2.2.3 Bias methods validation phase
In this phase, evaluation is first done for each site separately. Every site can be associated with a
nominal signal level. The assumption is that this signal level will be located somewhere around
the median level: (
) .
Defective pixels are assumed to deviate significantly from this nominal level, similar to the
discussion in 4.3.1.2.2. From the perspective of bad pixel detection however, a pixel is only
flagged as ‘bad’ when a clear distinction can be made. When unsure about a pixel’s behaviour, it
will be flagged as ‘suspect’. Hence, it is more likely that in this phase, a lot of pixels will be
flagged ‘suspect’ instead of ‘bad’.
When all sites have been evaluated, a pixel status list is available for each of them. These status
lists are cross-validated as follows:
if a pixel is flagged as ‘bad’ in one site, and is flagged as ‘bad’ or ‘suspect’ in all other
sites of a later date, it is considered as ‘bad’ over all these sites. Thus, one indication is
good for all.
however, if a pixel is flagged as ‘bad’ in one site, and is flagged as ‘good’ in a site at a
later date, this indicates a misdetection. The method is then invalidated, and the
evaluation is re-checked manually for this pixel.
if a pixel is flagged as ‘suspect’ in one site, and is flagged ‘suspect’ or ‘good’ in all other
sites (of a later date), it is considered as ‘suspect’ over all these sites.
if a pixel is flagged as ‘good’ for all sites, it is considered a good pixel and does not
undergo further processing.
4.3.2.2.4 Bias methods integration phase
When the validation phase is finished, a final result for the pixel status list is achieved. It is then
checked if the information of some sites can be considered redundant wrt others. This is checked
as follows:
when one site allows for all detections of suspect or aberrant pixels as another site, and
allows for more detections still, then this site is superior and the other site is removed
from the evaluation list.
When all sites have been checked in this manner and redundant sites have been removed, the
remaining sites are then integrated in one system. This system is then tested with new data: if the
system produces the same pixel status list result for 10 subsequent trials as the full list of sites, the
system is considered validated. It is expected no further trials are necessary, considering the
validation is done over all pixels.
4.3.2.2.5 Gain method phase
In the gain method, only the pixels flagged as ‘suspect’ are still under evaluation.
When moving to the gain method, averaged pixel values from different sites need to be
compared. Since the site data is possibly captured under different imaging conditions, this needs
to be compensated for first. Hence, the differences are rescaled with the nominal site value as
defined above : k
meangsite
kmeangik
meanscalediDN
DNDN
,,
,,,, .
When the gain is defective, the value for higher signals (ie. brighter sites or sites with stronger
gain) deviates more from the nominal value than the value for lower signals does. Therefore, the
detection algorithm proceeds as follows:
sort site values from low to high values
calculate for each site the deviation of pixel values from nominal values:
1,,, k
meanscaledi
k
sitei DNDN
when this deviation increases significantly for high values, the pixel is flagged as ‘bad’.
If it doesn’t, no further indication of a defect is possible, and the pixel status remains
‘suspect’.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 32/139
4.3.2.2.6 Full evaluation phase
When all methods are finished, the full chain of methods is evaluated. As the goal of the defect
detection is to flag pixels either as ‘bad’ or ‘good’, the pixels still flagged as ‘suspect’ at this
stage are held under scrutiny. When it is assured that no better conclusion can be made based on
the existing data, the pixel’s flag is set to ‘good’. This is the only acceptable conclusion: for the
time being no defective behaviour can be observed in the pixel. As a final output, a pixel quality
map is therefore made showing pixels either as ‘good’ or ‘bad’.
In the full evaluation phase, the defect detection procedure can be considered to be finalized.
Some further comments need to be made then:
when a pixel is flagged as ‘bad’ in the final procedure, it is considered defect for all
further dates. Therefore, it makes no sense to re-investigate its state in a later detection
procedure. The pixel quality map is therefore handled as an input to eliminate ‘bad’
pixels at the start of the procedure
also, the uniform site data used in the procedure is likely to be used by other calibration
operations as well. The dependencies between these must be clearly defined.
Finally, the statistics used in the detection methods are controlled by thresholding values
which can be controlled by the user. These, together with the finalized pixel status list,
should be transparent to the calibration manager. Table 5 shows a full list of the
thresholding values.
Parameters
Used in:
Line
outlier
detection
Bias
deviation
calculation
Gain
deviation
calculation
REQ_LINES
single
value
k
thresholdSTD, value per
band
k
lowthresholdsiteDEV _,
value per
band and
site
k
highthresholdsiteDEV _,
value per
band and
site
single
value
single
value
value per
band
Table 5: Thresholding values for defect detection
4.3.2.3 PF automatic operations
An automatic pixel detection method is described but not implemented in the PF DPM for the
analysis of SWIR pixels. A SWIR sensor tends to be quite sensitive to radiation. For this reason,
the PF processing might be disadvantaged if it must wait for an update of IQC calibration files
before knowing which pixels have become defective since the previous calibration file update. A
quick automatic pixel detection allows the PF to assess obvious cases of new bad pixels
independently before the arrival of new IQC calibration files.
The PF detection method is based on the study of signal stability, and of nearest neighbour
deviations. Signal stability is first checked: when the signal of a pixel is very unstable over a
certain number of lines, it is considered as an aberrant pixel. Stable signals are then averaged, to
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 33/139
assess the average value of a pixel. These are then checked with nearest neighbours, according to
a voting strategy. Very high deviations are a mark of aberrant behaviour, significant but not so
high deviations are still considered as a correctable offset (similar to a dark current). The full
motivation of this method is described in [PVDOC-624].
4.3.3 Linearity Check
4.3.3.1 Introduction
The measurement of the linearity of the relation between effective spectral radiance and digital
output is crucial, as systematic deviations from this linearity are able to happen when in-flight.
Possible causes of this are:
saturation of the sensor because of surface full well (interface traps capturing electrons)
saturation of the electronics because of voltage cut off, when an input has a higher
voltage than the maximum voltage corresponding to the maximum bit being registered.
A reliable method of measuring linearity is by changing the integration time, while imaging a
stable light source. In the assumption that the input is stable and integration time always has a
linear relationship, the dynamic range can be explored by collecting more or less signal in
varying the integration time. This process is illustrated in Figure 4.
Figure 4: Measurement of linearity, by varying integration time.
The incoming radiance generates a current which is collected during the configured integration
time. At large integration times, more charge is collected, and saturation defects are more likely
to occur. By varying integration time, it can be assessed at which integration time (and thus, at
which collected charge) saturation defects become significant. An integration time setting that is
used operationally must be below this threshold integration time. If not, the response is nonlinear,
without the sensor model being able to correct for it.
In the context of vicarious calibration, by a stable light source is meant a uniform calibration
zone. The signal magnitude over this zone should be low enough to be still able to measure the
response at the range of integration times to be tested (if signal magnitude is too high, saturation
will be reached too soon). Ideal candidates for these zones are the desert zones that are used for
multi-temporal calibration operations (see section 4.5.3).
4.3.3.2 Algorithm
4.3.3.2.1 Establish radiance outputs
input:
toperational
L
I tint
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 34/139
Images captured over deserts with a set of integration times, spread over the dynamic
range of the instrument for different bands: (set TBD).
Sensor model parameters: A, g, dc. All the parameter sets for the bands, pixels,
temperatures, gain numbers under investigation need to be defined.
procedure:
For each of the image points, remove all lines that are outliers, and determine the average
result over each different integration time set for each pixel using the procedure described
in section 4.3.2.2.1.
Derive the effective spectral radiance L for each of the image points and associated
parameter sets, using Equation 4.
For each of the image points, collect the L values for the different integration time sets. If
the linearity degrades, the values will show a similar behaviour as depicted in Figure 5. A
more clear distinction can be made when the values are multiplied with their respective
integration time, so values L x tint are investigated in function of tint.
Apply linear regression to the ensemble of data points. If no significant regression error is
obtained, the entire ensemble is in the linear regime. If a significant error is obtained,
proceed to the next step.
Apply linear spline fitting to the ensemble of data points. It is expected that two or three
regions will be found: one regime describing nominal levels, one regime describing
saturation, one (optional) regime describing low level deviations.
Figure 5: Illustration of linearity degradation.
output:
The regression error corresponding to the linear or linear spline model that is obtained.
If not sufficiently linear, the coefficients of the linear spline are also provided.
4.4 Absolute radiometric calibration
4.4.1 Rayleigh calibration: oceans
4.4.1.1 Introduction
The aim of the so-called Rayleigh calibration is to determine the absolute calibration coefficient
Ak
for the BLUE and RED bands. This method has been successfully applied to other sensors
(POLDER, PARASOL, SPOT/VEGETATION, MERIS,…) and detailed descriptions exist in the
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 35/139
literature (Vermote et al., 1992 [LIT2 ]; Hagolle et al., 1999 [LIT3]; Fougnie et al., 2007 [LIT4];
Martiny et al., 2005 [LIT5]).
The at-sensor radiance at short wavelengths over dark deep oceans under large solar and viewing
angles comes mainly from Rayleigh scattering (scattering by air molecules). Rayleigh or
molecular scattering can accurately be calculated based on the surface pressure and viewing
angles. The contribution of aerosol scattering (more specifically the aerosol optical thickness
(AOT), can be derived from the NIR reference band where molecular scattering is very low (due
to the strong wavelength dependence of Rayleigh scattering of -4
). The aerosol content estimated
from the NIR band is then transferred to the BLUE and RED band to model the TOA radiance
with a radiative transfer software module (6SV, [LIT6]). The simulated radiance values are then
compared with the measured values to derive the absolute calibration coefficient.
To reduce the perturbing part of the signal due to ocean reflectance and presence of foam, strict
pixel selection procedures are used. Pixels can only be chosen within oligotrophic areas with well
known weak and stable chlorophyll content. To minimize foam radiance meteorological data are
used to select zones with low wind speed. Finally only pixels outside sun glint direction are used.
The Rayleigh calibration approach allows for absolute calibration of the BLUE, and RED bands
taken the NIR band as reference band. The method cannot be applied to the SWIR band. The
results can be transferred to other bands (NIR, SWIR) based on inter-band calibration approaches
as sun glint (for both NIR and SWIR) or clouds (NIR) as described in section 4.5.1 and 4.5.2
respectively.
4.4.1.2 Algorithm
4.4.1.2.1 Physics of the algorithm
4.4.1.2.1.1 Conversion of TOA radiance to TOA reflectance
Following the inverse sensor model relation (Equation 4), the radiance k
TOAL at the TOA for pixel
i is derived from the measured digital number DN. k
TOAL is then normalized by the exo-atmospheric solar incident flux in order to obtain the TOA
reflectance k
TOA :
2
00 )cos(
dd
E
L
s
k
k
TOAk
TOA
Equation 8: Conversion of TOA radiance to TOA reflectance
with kE0 is the mean extra-terrestrial solar irradiance ( CEOS recommended irradiance file :
Thuillier ,2003 [LIT7] ) integrated over the spectral response of the different PROBA-V
spectral bands (Sk()) as
0
0
0
0
dS
dES
E
k
k
k
s is the solar zenith angle
dd0 is the ratio of Sun-Earth distance at the acquisition date to the mean Sun-Earth
distance, calculated based on the Julian day according to the formula in the 6SV subroutine
varsol.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 36/139
4.4.1.2.1.2 TOA reflectance signal decomposition
Over oceans the TOA reflectance TOA (outside strong oxygen and water vapour absorption
bands) can be decomposed as (omitting spectral band index and angles for simplicity):
w
vtotstotw
pathgasTOAs
TrTrTr
1
,,
Equation 9:TOA reflectance decomposition
with
w the water leaving reflectance
stotTr , , vtotTr , the total atmospheric transmittance for aerosols and molecules (sum of
direct and diffuse components) respectively for the solar and viewing
zenith angles
s spherical albedo of the atmosphere
gasTr the total gaseous transmittance (assumed to be decoupled from the
scattering)
path the path reflectance due to molecular and aerosol scattering in the
atmosphere, including coupling terms and the specular reflection by the
wavy surface
We ignore here the surface foam reflectance, which is negligible for surface wind speeds under
10 m/s (wind speed mask applied in processing). The specular reflection of direct sun light or the
sun glint reflectance is also ignored as for Rayleigh calibration only pixels outside sun glint
direction are used (glitter mask applied in processing).
The path reflectance path
ρ can be decomposed into the following components :
gaagmmpath
Equation 10: Path reflectance decomposition
with
m reflectance due to multiple scattering of air molecules
(Rayleigh reflectance) in absence of aerosols
gm reflectance due to specular reflection at the rough sea
surface of the light scattering by the molecules in the atmosphere
a reflectance due to multiple scattering by aerosols and by combined
successive scattering by molecules and aerosols; This term is computed
as mmaa
.
ga reflectance due to specular reflection at the rough sea surface of the
light scattering by combined successive scattering by molecules
and aerosols
Because the Fresnel surface reflectivities are small, the additional path reflectance terms gm and
ga are smaller than the direct path reflectance terms ( m ,a ). Because wind speed alters the
magnitude of the reflected radiance, gm and ga depend on the wind speed.
The water leaving reflectance in the near infrared (NIR) wavelength region is generally equal to
zero for dark oceanic waters Gordon and Wang (1994) [LIT8] due to the strong seawater
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 37/139
absorption. This is the so called black pixel assumption. Therefore the equation for the TOA
reflectance (Equation 9) for the NIR band can be simplified as
)()()( NIRNIRTrNIR pathgasTOA
Equation 11: NIR TOA reflectance simplified decomposition
In Figure 6 the preliminary spectral response curves of PROBA-V bands are given with respect to
the major gaseous absorption bands. These figures indicate that gaseous transmittance in the
PROBA-V NIR band is (slightly) contaminated by water vapour. For dark surfaces like oceans
we can however not fully decouple water vapour absorption and scattering (Martiny et al, 2005
[LIT5]) because aerosol and water vapour can be located at the same level in the atmosphere. As
information on the relative scale height of the water vapour and of the aerosols is generally not
available, we will adopt the scheme given by Vermote et al. [LIT6]in the 6SV manual. This
scheme assumes that the aerosol path radiance is generated above the middle of the boundary
layer. Thus the additional attenuation is made by half the precipitable water:
)(,
2,,,)(
2
222
23 gagmOH
OH
a
OHOH
mTOA UMTrU
MTrMOOTrNIR
with
M the air mass given by vs
M cos
1
cos
1
MOOTr ,, 23 the gaseous transmittance of
23 ,OO
OHTr 2 the water vapour transmittance.
OHU2
the total precipitable water in units of g/cm²
The TOA apparent reflectance corrected for the gaseous transmittance of 23 ,OO will be denoted
as :
),,( 23 MOOTr
TOAc
TOA
Equation 12: Correction of TOA reflectance for gaseous (Ozone, Oxygen) absorption
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 38/139
Figure 6: Gaseous absorption in PROBA-V bands
4.4.1.2.1.3 Retrieval of the aerosol optical thickness (AOT) from NIR band
The gaseous transmittance and molecular scattering can be calculated with 6SV based on the
available ancillary data (surface pressure, amount of ozone and total precipitable water, wind
speed). 6SV includes polarization and specular reflection by the wavy surface. The only unknown
in Equation 11 is the aerosol scattering contribution, which depends on the aerosol optical
thickness (AOT) and the aerosol model.
The aerosol model adopted for the 6SV calculations is the Shettle and Fenn Maritime aerosol
model with 98% humidity (denoted as M98) (Shettle and Fenn, 1979 [LIT9]). Following Hagolle
et al. (1999) [LIT3] highest accuracies can be achieved with the M98 model (errors due to
deviation from M98 are taken into account in the error budget). The maritime aerosol model has
two components (a) a sea-salt component and (b) a continental component which is assumed to
be identical to the rural aerosol with the exception that very large particles were eliminated. The
M98 aerosol model can be introduced in 6SV by specifying the particle size distribution of each
component of the aerosol model, represented by a log normal distribution function. Tables with
individual component refractive indices (real and imaginary) in function of humidity versus
wavelength are given in the report by Shettle and Fenn (1979) [LIT9].
The aerosol optical thickness (AOT) is however not a priori known and will be derived from the
NIR band. The primary assumption of Gordon and Wang (1994) [LIT8] is that the ocean (i.e. the
water leaving reflectance) does not contribute to the TOA signal in the NIR, (i.e., all radiance
reaching the sensor is of atmospheric origin (including fresnel reflection on the ocean surface)).
H20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
wavelength
H20 t
ran
sm
itta
nce
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
PR
OB
A-V
Sp
ectr
al
resp
on
se c
urv
es (
TB
C)
O3
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
wavelength
O3
tra
ns
mit
tan
ce
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
PR
OB
A-V
Sp
ectr
al
resp
on
se c
urv
es (
TB
C)
O2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
wavelength
O2
tra
ns
mit
tan
ce
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
PR
OB
A-V
Sp
ectr
al
resp
on
se c
urv
es (
TB
C)
CO2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
wavelength
CO
2 t
ran
sm
itta
nc
e
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
PR
OB
A-V
Sp
ectr
al
resp
on
se c
urv
es (
TB
C)
CH4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
wavelength
CH
4 t
ran
sm
itta
nc
e
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
PR
OB
A-V
Sp
ectr
al re
sp
on
se
cu
rves (
TB
C)
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 39/139
Using this assumption, the NIR bands can then be employed for the aerosol determination in the
calibration process. This assumption holds for Case 1 waters with low chlorophyll concentration
and where phytoplankton is the only optically significant water column contributor.
The AOT per pixel is retrieved from the NIR band based on a LUT. This LUT contains modelled
)(mod, NIRelc
TOA in function of geometrical parameters, pressure, water vapour, wind speed and
AOT. The LUT of simulated/modelled )(mod, NIRelc
TOA are interpolated (on the angles, water
vapor, windspeed, pressure of each pixel) to obtain a LUT of NIRelc
TOA
,mod, in function of only
AOT. The AOT value for which the measured )(, NIRmeasc
TOA fits best with the modelled
)(mod, NIRelc
TOA is searched for.
4.4.1.2.1.4 Retrieval of the calibration coefficients for the BLUE and RED bands
Once the AOT is determined )(mod, REDelc
TOA and )(mod, BLUEelc
TOA can be calculated according
to Equation 9 with the radiative transfer model 6SV. The water leaving reflectance in 6SV is
calculated using the semi-analytical reflectance model described in Morel, 1988 [LIT10]. This
model is based on the bulk diffuse attenuation coefficient or downward irradiance . On
average this coefficient depends on the chlorophyll content (CHL) in Case1 waters, according to
= .
Where corresponds to the contribution of pure water to ; the coefficient and
exponent ) are tabulated values obtained through regression analysis performed on and CHL quantities in Case 1 waters. The equation accounts implicitly for coloured
dissolved organic matter (CDOM) associated and co-varying with the CHL concentration in
Case 1 waters. The are transformed in water leaving reflectance wρ in an iterative way.
The water leaving reflectance wρ in Equation 9 is therefore a function of the CHL the
chlorophyll content.
The Rayleight calibration is based on pre-defined oligotrophic waters with low and very stable
chlorophyll concentrations. The calibration zones (Table 6) are taken from Fougnie et al.,2002
[LIT11] who selected these zones based on stability and uniformity criteria during two years
(1998, 1999) of SeaWIFS level 3 products.
Table 6: Stable oligotrophic oceanic sites
In the recent studies by Fougnie et al. (2010) [LIT56] the monthly variation of marine reflectance
for the defined calibration zones is analyzed on the basis of 9 years of SEAWIFS data. A similar
study has been performed by Morel et al (2010)[LIT54]. In this latter study, the monthly averaged
CHL and CDOM concentrations for six selected oligotrophic areas (similar to or located within
the zones listed in Table 6) are reported. As shown in Figure 7 notable difference in the seasonal
variation of CHL occur. The seasonal variation is less pronounced and featureless in the North-
West of Pacific and in the The North Pacific gyre. For this latter the highest annual mean CHL
(0.059 mg m-3
) are observed. The seasonal CHL pattern for the other 4 oligotrophic zones are
comparable and rather reproducible from year to year with peaks occurring around July-August.
Name Min. latitude (deg) Max. latitude (deg) Min. longitude (deg) Max. longitude (deg)
South of Atlantic -19.9 -9.9 -32.5 -11
South of Indian -29.9 -21.2 89.5 100.1
North of Atlantic 17 27 -62.5 -44.2
North of Pacific 15 23.5 179.4 200.6
North west of Pacific 10 22.7 139.5 165.6
South-East of Pacific -44.9 -20.7 -130.2 -89
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 40/139
Figure 7: Monthly averaged chlorophyll (in mg m-3
) for each Rayleigh calibration zone
To model )(mod,
REDelc
TOA and )(
mod,BLUE
elcTOA
with 6SV the montly varying and site-
depended CHL values as given in Figure 7 are used. The coefficient and exponent ) are
replaced with the slightly modified values from Morel et al (2007) [LIT55].
The Morel (1988) [LIT10] semi-analytical model used in 6SV implicitly includes a “mean”
relationship between CHL and CDOM. In recent studies by Morel et al (2010) [LIT56 ]and
Morel and Gentili (2009)[LIT53] deviations from these mean relationship has been observed and
studies for the different oligotrophic zones. To account for these difference in the Morel (1988)
[LIT10] semi-analytical model, the model could be operated with geographically distinct values if avalaible for these zones (at this moment only published for the South Pacific).
The evaluation of the calibration coefficients consists in comparing )(mod, REDelc
TOA and
)(mod, BLUEelc
TOA computed with the radiative transfer model 6SV (model) with )(, REDmeasc
TOA
and )(, BLUEmeasc
TOA derived from the PROBA-V sensor measurements (meas), assuming a given
(old) calibration (the one we want to evaluate).
Consequently, the ratio kA , relative change in calibration coefficients, defined as:
)(
)(mod,
,
RED
RED
A
AA
elc
TOA
measc
TOA
old
RED
new
REDRED
and
)(
)(mod,
,
BLUE
BLUE
A
AA
elc
TOA
measc
TOA
old
BLUE
new
BLUEBLUE
provides a measure of the difference in calibration coefficients with respect to the reference
values (old). This ratio will be determined for each selected pixel. In section 0 it is explained
how these individual estimates of newA from the different pixels, sites and acquisition days will
be statistically analyzed to define the final newA .
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 41/139
4.4.1.2.2 Algorithm implementation
4.4.1.2.2.1 Generation of LUTs
As it is computationally prohibitive to run a radiative transfer model for every pixel of all the
Rayleigh calibration images, a LUT approach will be used. The LUT will be created using 6SV
(Vermote, 2008 [LIT6]) simulations. As its scalar predecessors, the vector version of 6S is based
on the method of successive orders of scattering (SOS). The effects of polarization are taken into
account in the vector version. Accounting for radiation polarization is extremely important for
radiative transfer calculations over dark targets such as ocean surfaces. Ignoring the effects of
polarization leads to large errors in calculated top-of-atmosphere reflectances: more than 10% for
a molecular atmosphere and up to 5% for an aerosol atmosphere (Kotchenova et al., 2007
[LIT12]).
For the NIR band a LUT will be created for apparent TOA reflectance, corrected for gaseous
transmittance of 2O,3O for an atmosphere bounded by black ( wρ =0) fresnel reflecting wind
roughened oceanic surface. This LUT will be a function of sun zenith angle, view zenith angle,
relative azimuth angle, wind speed, pressure, water vapour and AOT.
Often LUTs are only calculated for standard atmospheric pressure of 1013.25 hPa and correction
for atmospheric pressure variation are performed using the empirically retrieved formula from
Gordon et al. (1988) (LIT13). Wang (2005) [LIT14] has however showed that this approach
result in an extra uncertainty, especially in the blue. To improve the performance of the Rayleigh
calibration approach we will use the Rayleigh solution of the radiative transfer calculations as
done with 6SV. The effect of surface pressure variation can be exactly calculated with 6SV by
changing zP in the atmospheric profile (Kotchenova, private communication, July 2009).
For the BLUE and RED bands also LUTs are created for the apparent TOA reflectance,
corrected for gaseous transmittance of 23 ,OO for an atmosphere bounded by fresnel reflecting
wind roughened oceanic surface with a chlorophyll concentration taken from
in function of the month and site. These LUTs are a function of sun zenith angle, view zenith
angle, relative azimuth angle, wind speed, pressure, water vapor and AOT.
For the generation of the LUTs with 6SV two options are considered: (a) a global pre-calculated
LUT and (b) an on-the-fly calculated LUT. The global LUT will be valid for all imagery and will
contain cTOAρ for a variety of sun-view geometries, AOTs, pressures, water vapor concentrations
and wind speeds within certain ranges. The on-the-fly calculated LUT is much smaller and only
valid for the image under consideration. The ranges of the parameters are constrained by the
ancillary image and meteorological data.
4.4.1.2.2.2 Processing on the PROBA-V calibration imagery
Conversion to meas
TOA : The input DNs for all image pixels are converted to TOA reflectance k
TOA
following Equation 4 and Equation 8.
Clear pixel selection : Only pixels at a distance of at least 30 km from clouds will be kept, by
extending the standard cloud mask. Furthermore only pixels for which the retrieved AOT in the
NIR band is less than 0.05 will be selected in order to reject data with sensible aerosol loading.
Masking white caps : To avoid contamination by white caps or foam only pixels for which the
wind speed (from the meteorological data) is smaller than 5 m/s will be kept.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 42/139
Masking pixels probably affected by sun glint : To avoid sun glint only pixels with n > 20° are
selected:
2cos2
coscoscos
p
vs
n ar
with cossinsincoscoscos vsvsp ar
Correction for gaseous absorption of ozone and oxygen; conversion from meas
TOA to measc
TOA
, :
The correction is made using a pre-calculated exponential variation with air mass and gaseous
amount, for example
nUgairmssaMgasTr )^*(*exp,
Equation 13: Calculation of gaseous transmittance
where Ug is the gaseous concentration (eg. total amount ozone in units of cm/atm; the amount of
oxygen is calculated via pressure), a and n are coefficients which depend on the response of the
given spectral band. Details can be found in Tanré et al (1990) [LIT15] and the SMAC model
(Rahman and Dedieu, 1994 [LIT16]). The 6SV radiative transfer model is used to calculate
MOOTr ,, 23 in each PROBA-V spectral band for a range of total amount of gas and a view
angles. These 6SV runs are used to derive the coefficients of the exponential relationship.
Estimate for each pixel the aerosol optical thickness (AOTest
): the simulated )(mod, NIRelc
TOA (in
the global LUT or the on-the-fly calculated LUT) are interpolated on the angles, water vapour and
pressure from each pixel to obtain )(mod, NIRelc
TOA corresponding to the geometry and atmospheric
conditions (water, pressure,windspeed) of the observation. Then ),(mod,
i
elc
TOA AOTNIR and
),( 1
mod,
i
elc
TOA AOTNIRthat surround
)(, NIRmeasc
TOA are the aerosol optical thickness (AOT
est)
corresponding to )(, NIRmeasc
TOA is estimated through interpolation.
Calculate BLUEA and REDA :
For each pixel find ),(mod, REDBlueelc
TOA corresponding to AOTest
: ),,(mod, estelc
TOA AOTREDBlue
is derived from interpolation between ),,(mod,
i
elc
TOA AOTREDBlue and
),,( 1
mod,
i
elc
TOA AOTNIRBlue at the estimated AOTest
. Next, for each pixel the change in
calibration coefficients for the BLUE and RED bands is calculated as :
)(
)(mod,
,
BLUE
BLUE
A
AA
elc
TOA
measc
TOA
old
BLUE
new
BLUEBLUE
;
)(
)(mod,
,
RED
RED
A
AA
elc
TOA
measc
TOA
old
RED
new
REDRED
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 43/139
4.4.1.2.2.3 Required ancillary data
Ancillary data required for the pixel selection, gaseous transmittance correction and LUT
interpolation are:
Wind speed at the sea level
This parameter can be accessed through the European Center for Meteorological and Weather
Forecast (ECMWF) . The typical accuracy attached to wind speed is of about 2 m/s.
Atmospheric pressure at the sea level
This parameter can be accessed through the European Center for Meteorological and Weather
Forecast (ECMWF). About 5-10 mbars is a reasonable accuracy for the atmospheric pressure.
Total ozone amount
This parameter can be accessed through the European Center for Meteorological and Weather
Forecast (ECMWF) or from the measurements of the Total Ozone Mapping Spectrometer
(TOMS). An accuracy of about 10-20 mAtm-cm (Dobson Units, DU) is a typical accuracy for the
total ozone amount of the atmosphere.
Water vapor content
This parameter can be accessed through the European Center for Meteorological and Weather
Forecast (ECMWF). An uncertainty of 20 % is considered.
Other ancillary data (attached to the input image) : time, date, sun zenith angle, solar zenith angle,
relative azimuth angle and a cloud mask (calculated from the image itself)
4.4.1.2.2.4 Summary of the processing steps
Figure 8: Overview flowchart Rayleigh calibration
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 44/139
4.4.1.3 Error analysis
The considered error sources are :
1) Surface pressure: a maximal error of 10 hPa is used in the error analysis
2) Wind speed: The wind speed impacts the contribution of photons scattered by the
atmosphere after their reflection over the sea-surface. A typical accuracy of 2 m/s in wind
speed is used in making up the error budget.
3) NIR calibration: an error of 3% is used to make the error budget
4) Aerosol model: The aerosol model is assumed fixed (M98); Uncertainties in the calibration
method related to the aerosol type are analyzed by calibrating a scene generated with a
coastal aerosol (C70) with a LUT generated with M98
5) Gaseous transmittance: An error in the ozone amount and water vapor of respectively 5 %
and 20 % is used in the error budget. An uncertainty on water vapour will have a small
direct effect on the RED band, and an indirect impact due to the aerosol retrieval in the
NIR band.
6) Chlorophyll content : an error of 50% on the chlorophyll concentration has been used for
making up the error budget. This will have a direct effect on the BLUE and RED band,
however as the water surface can be assumed totally black in the NIR band, an error in the
chlorophyll content will have no effect on the AOT retrieval.
The error analysis is performed by using a LUT generated according to the reference
specifications. This LUT is then used to perform a Rayleigh calibration on an artificial image
generated according to the same specifications except for the parameter of interest (e.g. C70
model instead of M98 model in LUT) or by introducing an error in the meteo-data (e.g. 20 %
error in water vapour).
Table 7 summarizes the error budget for the Rayleigh method for both BLUE and RED band. The
table contains the average error, the error at 1-sigma and at 2-sigma The error sources are
assumed to be non-correlated and thus the total error is the quadratic sum of all errors. The
impact of uncertainties related to the input parameters is estimated on the calibration result. As
uncertain impact parameters were considered : pressure, wind speed, NIR calibration, aerosol
model, gaseous amount and marine reflectance (chlorophyll concentration).
As the exact spectral response curves for the PROBA-V bands are not yet known, this error
analysis may be updated in a later phase.
Error Sources
BLUE RED
average
error
(%)
1-sigma
error
(%)
2-sigma
error
(%)
average
error
(%)
1-sigma
error
(%)
2-sigma
error
(%)
Pressure 0.571 0.632 0.693 0.364 0.394 0.425
Wind speed 0.570 1.535 2.501 0.735 2.035 3.335
NIR calibration 0.580 0.650 0.720 1.822 1.943 2.065
Aerosol model 0.536 0.959 1.331 1.070 1.825 2.580
Ozone 0.029 0.032 0.036 0.267 0.300 0.333
Water vapour 0.047 0.102 0.157 0.089 0.135 0.181
Chlorophyll 0.951 1.106 1.261 0.280 0.334 0.388
Total absolute 1.477 2.310 3.262 2.301 3.409 4.745
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 45/139
Table 7: Summary of Rayleigh absolute calibration error budget.
In the BLUE band the performance is determined by the accuracy on the wind speed, marine
reflectance and aerosol model. In the RED band the error in the NIR calibration, wind speed and
aerosol model are the limiting factors to accuracy. In section 0 it is explained how the uncertainty
calculation is used in the determining the final error budget.
4.4.2 Reflectance-based method
4.4.2.1 Introduction
In the reflectance-based method ground-based reflectance measurements are used as input. At the
time of the PROBA-V overpass a field crew measures the spectral reflectance of the selected
calibration site using a well-calibrated field spectroradiometer under optimal (cloud-free) weather
conditions. Furthermore, the atmospheric and meteorological conditions during the overpass have
to be accurately characterized using sun photometer measurements and meteo station data. Using
a radiative transfer model the surface reflectance spectra can be converted to TOA radiance and
compared to the measured PROBA-V radiance. (The accuracy of the reflectance-based method is
generally lower than for the radiance-based method due to uncertainties in the aerosol modelling.)
A higher accuracy is possible by the improved reflectance-based method which uses the same
measured data as from the reflectance-based method as well as the measured ratio of diffuse-
global spectral irradiance at ground level. Because of the mobilization costs of the field crew, the
reflectance-based method will only be applied to validate the operational calibration methods at
an ad hoc basis through collaboration with other teams or by joining calibration/validation
exercises like the CEOS 2010 campaign at Tuz Gölü, Turkey, one of the eight CEOS endorsed
core instrumented IVOS Sites (Table 8) provisionally called LANDNET sites. The characteristics
of these eight LANDNET sites can be obtained from
http://calval.cr.usgs.gov/sites_catalog_ceos_sites.php.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 46/139
Core
“Instrumented”
IVOS Site
Center
latitude
Center
longitude
Point of Contact
Railroad Valley
Playa, NV,
USA, North
America
38.50 -115.69
Dr. Kurtis J. Thome ([email protected]) –
NASA
Ivanpah,
NV/CA, USA,
North America
35.57 -115.40
Dr. Kurtis J. Thome ([email protected]) –
NASA
Lspec
Frenchman Flat,
NV, USA, North
America
36.81 -115.93
Carol J. Bruegge ([email protected]) –
NASA JPL
La Crau, France,
Europe
43.56 4.86 Patrice Henry ([email protected]) – CNES,
France
Dunhuang, Gobi
Desert, Gansu
Province, China,
Asia
40.13 94.34
Xiuqing Hu ([email protected])-NSMC/CMA, China
Negev, Southern
Israel, Asia
30.11 35.01 Arnon Karnieli ([email protected]) – Ben Gurion
University, Israël
Tuz Gölü,
Central
Anatolia,
Turkey, Asia
38.83 33.33
Selime Gurol ([email protected]) –
TUBITAK UZAY, Turkey
Dome C,
Antartica
-74.50 123 Stephen Warren ([email protected]) –
University of Washington, USA
Table 8: CEOS Core Instrumented IVOS Sites (LANDNET sites).
These instrumented reference test sites are primarily used for field campaigns to estimate the
radiometric absolute calibration coefficients and can serve as a focus for international efforts,
facilitating traceability and cross-comparison to evaluate biases of in-flight and future sensors in a
harmonized manner.
An in-flight radiometric calibration experiment needs to be performed under clear sky conditions
above at least one of the prime large and homogenous earth calibration targets (Teillet et. al.,
2007 [LIT17]). In case of non-linearity of PROBA-V detectors, more calibration sites should be
selected covering the full dynamic range (water, sand, ice,…).
Radiometry reference test site selection criterion put forward by the CEOS WGCV subgroup
IVOS are listed in Table 9 (Teillet et al., 2007 [LIT17]).
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 47/139
Selection criterion Purpose
High spatial uniformity over a large area (within
3%)
Minimize misregistration and adjacency effects
Surface reflectance greater than 0.3 Provide higher SNR and reduce uncertainty due to
atmosphere
Flat spectral reflectance
Reduce uncertainties due to different spectral
response profiles
Temporally invariant surface properties (within
2%)
To reduce BRDF, spectral, surface reflectance
effects
Horizontal surface with nearly lambertian
reflectance
Minimize uncertainty due to different solar
illumination and observation geometry, to
minimize slope-aspect effects
At high altitude, far from ocean, urban, and
industrial areas
Minimize aerosol loading and atmospheric water
vapour, to minimize anthropogenic aerosols
In arid regions with low probability of cloud cover Minimize precipitation that could change soil
moisture and to increase the probabilityof the
satellite instruments imaging the test site at the
time of the overpass.
Table 9: Radiometry reference test site selection criterion.
An on-line Catalogue of worldwide test site for sensor characterization is available at
http://calval.cr.usgs.gov/sites_catalog_map.php.
CEOS envisages a free and open data policy. Satellite products and in-situ data from a number of
test sites (not all LANDNET sites are included yet) can be queried free of charge from the CEOS
cal/val portal (http://calvalportal.ceos.org/cvp/web/guest). The objective of CEOS is to provide
by 2012 through the CEOS cal/val portal satellite and in-situ data from all LANDNET sites free
of charge to the cal/val community (N. Fox, private communication, December 2009).
Recently added to the list of CEOS instrumented site is the calibration site at Tuz Gölü, Turkey.
Tuz Gölü (Latitiude 38°46’21.00’’N, Longitude 33°28’23.16’’E ) is a salt lake which dries to a
salt surface for about 3 month a year (mainly halite and gypsum with minor amounts of polyhalite
and coelestine) in central Anatolia in Turkey far from the influence of the sea. In July and August
the dried salt lake has a homogeneous area over 324.026 km². The dried salt lake surface is
smooth, uniform (spatially uniform with RMS of deviation from mean smaller than 2% for a
300x100 m target), flat (slope less than 1°) and with high surface reflectance (0.4-0.58 in VNIR).
July and August are the optimum months for reflectance-based radiometric calibration as these
months are the most dry, sunny and cloud-free months of the year based on local data between
1987 and 2007. All these characteristics together with the fact that Tuz Gölü is easily accessible
especially for Europe makes this site well suited for the radiometric calibration of PROBA-V.
4.4.2.2 Algorithm
4.4.2.2.1 Physics of the algorithms
Reflectance-based method
The reflectance-based method uses ground-based reflectance measurements as primary input. At
the time of the overpass the upwelling radiance of the test site is measured and divided by the
measured radiance of the calibration panel to become the reflectance of the test site.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 48/139
Measurements should be taken in a short time frame (40 minutes) to prevent change of
atmospheric conditions and changing illumination conditions. Measurements are taken over
ground blocks of 100 m x 100 m (size of nadir PROBA-V BLUE, RED, NIR pixels) and
averaged to account for spatial variation. The size of the measurement site is typically 4 cross
track pixels and 16 along track pixels (Biggar et al., 2003 [LIT18]) which means for PROBA-V
16 x 4 100 m pixels. As described in Thome, 2001 [LIT19] measurements with a field
spectroradiometer are taken parallel to the across track direction in the center of the 4 across track
pixels. This sampling strategy is repeated 16 times for the along track pixels. With the field
spectroradiometer (8° FOV foreoptic at 2 m height means a ground diameter of 0.3 m) 30 spectra
per sample are collected. For each pixel 10 spectra are averaged. That means in total
300x4x16=19200 (N) spectra and 640 samples are collected over the total area of the site
(sampling strategy to be updated). Collecting this data takes 45 to 60 minutes. Reflectance is
determined by the ratio of the radiance of the reference target Lref by the BRDF corrected
radiance of the spectralon calibration panel LBRDF
panel for which the BRDF factor is determined in
the lab. The calibration panel radiance is measured at the start and end of the data collection and
after each 8 pixels or 80 samples.
pixels
BRDFref
panelL
Lref
N
1
Equation 14
All reflectance spectra are averaged to give one single reflectance spectrum for the measured test
site. Global, downwelling irradiance data are also collected near the test site to determine if there
are significant changes in diffuse skylight illumination during the measurement period.
Positions of the ground-based measurements need to be recorded by a Global Positioning System
(GPS) to locate the measured pixels in the PROBA-V image. Another possibility is to mark the
corners of the site with large tarpaulins visible in the PROBA-V image.
The atmospheric parameters like aerosol optical depth, Angström coefficient and water vapour
content is determined from sun photometer measurements following the Langley method based
on measurement of direct solar radiation whose voltage output V can be written as: M
gass eTrDVV 0
Equation 15
with V sun photometer voltage output, V0 calibration coefficient, Ds Sun-Earth distance factor
given by
)4(9856.0cos01673.01
1
JDDs
Equation 16
with JD the day of the year and Trgas gaseous transmittance, the total atmospheric optical depth
(sum of optical depth due to aerosol scattering and optical depth due to Rayleigh scattering) and
M air mass which depends on the solar zenith angle and the pressure.
When applying the natural logarithm on equation Equation 15, it can be written as:
Mlnln 0
V
TrD
V
gass
Equation 17
A plot of ln[V/DsTrgas] versus M for several solar zenith angles and an assumed stable
atmosphere gives V0 (=eintercept
) and the total optical depth (=-slope). From and Rayleigh (the
latter only depends on wavelength and pressure) aerosol can be determined for the bands not
affected by water vapour.
The spectral variation of aerosol optical depth can be written (according Angström’s turbidity
formula) as
)(aerosol
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 49/139
Equation 18
with the Angström turbidity coefficient that is proportional to the horizontal visibility (VIS)
according formula 15/VISe
Equation 19
and the Angström coefficient ( in function of wavelength).
By plotting measured aerosol in function of wavelength and fitting Equation 18 to these points,
(and thus VIS (in km)) and can be determined. Subsequently, knowing and , AOT550
which is equal to aerosol(550 nm) can be estimated from Equation 18.
By plotting the measured as a function of wavelength and fitting Equation 20 to these points,
at the water vapour affected band can be determined. /)( bae
Equation 20
The spectral band at 936 nm can be used to estimate water vapour content UW as there is an
important water vapour absorption band in this spectral region. Here the Trgas is not equal to 1 but
is estimated from 5175.05093.0 ..6767.0 MUW
gas eTr
Equation 21
The Langley equation can now be written as:
)6767.0(UWlnln 5175.00.50930
M
MVD
Ve
s
Equation 22
and thus V0 at 936 nm and UW (g/cm2) can be determined.
Using a radiative transfer model (such as 6SV) the measured reference reflectance refρ is
converted to TOA radiance refTOA
L using input aerosol optical thickness at 550 nm, Angström
coefficient, and water vapour determined from sun photometer measurements. Subsequently the
ground-based TOA radiance measurements are convolved with the PROBA-V spectral response
function S() according the following equation:
0
, / dSdSLL kk
ref
TOA
refkTOA
Equation 23
Finally refTOA
L are compared to the TOA radiance measured by PROBA-V measTOA
L (average of 4
along track x 16 along track pixels; we assume nadir PROBA-V viewing geometry, in case of
non-nadir viewing geometry a BRDF correction will be applied based on goniometer
measurements of the reference site).
This method can cause problems for sensors with coarse ground resolution (300 m-1200 m)
because of the required time to sample the test site. Another disadvantage are the assumptions
made for aerosol size and composition.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 50/139
4.4.2.2.2 Algorithm implementation
4.4.2.2.2.1 Processing
Ground-based measurements processing
Divide ground measured reference site radiance by BRDF corrected radiance of the
calibration panel to become ground-based reflectance (including a bidirectional reflectance
factor correction for the solar zenith angle)
Averaging of ground-based reflectance measurements
Radiative transfer calculations (e.g. 6SV) (with as input aerosol optical thickness at 550 nm,
Angström coefficient, and water vapour determined from sun photometer measurements) to
convert ground-based surface reflectance to TOA radiance refTOA
L .
Convolution with PROBA-V spectral response: the modelled refTOA
L data measurements are
convolved with the PROBA-V spectral bands.
PROBA-V processing
Extract calibration site pixels (4 across track x 16 along track pixels)
Conversion of DN to measTOA
L : The input DNs for all selected image pixels are converted to
TOA radiance measTOA
L following Equation 4.
Clear pixel selection : Visual inspection for clouds. The reflectance-based method is only
applied under optimal weather conditions (cloud-free).
Calculate kA : for each pixel the change in calibration coefficients for the BLUE, RED,
NIR and SWIR bands is calculated based on a comparison of PROBA-V and ground-based
TOA radiance as :
)(
)(
BLUEL
BLUEL
A
AA
ref
TOA
meas
TOA
old
BLUE
new
BLUEBLUE ;)(
)(
REDL
REDL
A
AA
ref
TOA
meas
TOA
old
RED
new
REDRED
)(
)(
NIRL
NIRL
A
AA
ref
TOA
meas
TOA
old
NIR
new
NIRNIR ; )(
)(
SWIRL
SWIRL
A
AA
ref
TOA
meas
TOA
old
SWIR
new
SWIRSWIR
4.4.2.2.2.2 Required ancillary data
Required ground-based equipment for determining ground-based TOA radiance
field spectroradiometer (e.g. FieldSpec Pro FR) for reflectance measurements
field spectroradiomete (e.g. FieldSpec Pro FR) + cosine receptor for downwelling irradiance
measurements
spectralon calibration panel
sun photometer
GPS
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 51/139
Ancillary data for the processing of the ground-based reflectance measurements to ground-based
TOA radiance
Bidirectional reflectance factor for the spectralon calibration panel as determined in the
lab
Aerosol optical thickness
This parameter is determined from sun photometer measurements and is input parameter of
the Radiative Transfer Codes (e.g. 6SV) to convert ground-based reflectance measurements
to TOA radiance.
Angström coefficient
This parameter is determined from sun photometer measurements and gives to a certain
extent an indication of size distribution and thus the aerosol type. From the Angstrom
coefficient the aerosol model is determined which is used in the Radiative Transfer Codes
(e.g. 6SV) to convert ground-based reflectance measurements to TOA radiance.
Water vapour content
This parameter is determined from sun photometer measurements and is input parameter of
the Radiative Transfer Codes (e.g. 6SV) to convert ground-based reflectance measurements
to TOA radiance.
Position
Position of ground-based reflectance measurements from hand-held GPS measurement to
localize the measurements in the PROBA-V image.
Terrain height
Terrain height (proxy for pressure) is an input parameter of Radiative Transfer Code (e.g.
6SV).
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 52/139
4.4.2.2.2.3 Summary of processing steps
Figure 9: Flowchart to calculate calibration coefficients
4.4.2.3 Error analysis
There are four basic areas of uncertainty in the method: (1) atmospheric characterization, (2)
surface characterization, (3) radiative transfer code, and (4) computation of the site-average DNs
(or radiance). The factors leading to uncertainties in determining the site average DN are an
incorrect determination of the site’s location in the image and the subsequent misregistration of
the site’s surface reflectance to DN (or radiance) (Thome 2001 [LIT19]).
The uncertainty has been estimated at 4.9% for the reflectance-based method and 3.5% for the
improved reflectance based method (Dinguirard and Slater, 1999 [LIT20]; Biggar et al. 1994
[LIT21]). With improved equipment and techniques soon an uncertainty of 3.3% and 2.8%
should become feasible.
Estimated measurement uncertainties for the reflectance-based and improved reflectance-based
methods are listed in Table 10 (Dinguirard and Slater, 1999 [LIT20]; Biggar et al. 1994
[LIT21]). The error given is the error in the TOA radiance caused by the error source. All
percentages are 1 . We assume the error sources are independent such that the total error is the
root sum of squares of the errors caused by all listed error sources.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 53/139
Error Sources Error
Ground reflectance measurement 2.1
Optical depth measurements 1.1
Absorption computations 1.3
Choice of aerosol complex index 2.0
Choice of aerosol size distribution 3.0
Vertical distribution 1.0
Non-lambertian ground characteristics 1.2
Non-polarization vs. polarization code 0.1
Inherent code accuracy 1.0
Uncertainty in the value of µsum 0.2
Total error (root sum of squares) 4.9
Table 10: Error sources of reflectance-based method.
4.4.3 Radiance-based method: APEX underflights
4.4.3.1 Introduction
APEX, Airborne Prism Experiment, is an airborne (dispersive pushbroom) imaging spectrometer
developed by a Swiss-Belgian consortium on behalf of ESA, (Itten et al., 2008 [LIT22]) to
calibrate/validate and simulate future spaceborne missions. The main APEX specifications are
listed in Table 11.
APEX main specifications
Spectral range VNIR: 380-970 nm
SWIR: 940-2500 nm
Spectral bands
VNIR: default 114 bands,
reprogrammable through
customized binning pattern.
Max. unbinned bands 334
SWIR: 199 bands
Spectral sampling interval
VNIR: 0.55-8 nm over spectral
range (unbinned)
SWIR: 5-10 nm over spectral range
Spectral resolution (FWHM)
VNIR: 0.6-6.3 nm over spectral
range (unbinned)
SWIR: 6.2-11 nm over spectral
range
Spatial pixels 1000
FOV (across track) 28°
IFOV 0.48 mrad
Spatial sampling interval (across track) 1.75 m @ 3500 m AGL
Sensor dynamic range VNIR: CCD, 14 bit encoding
SWIR: CMOS, 13 bit encoding
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 54/139
Pixel size VNIR : 22.5 µm x 22.5 µm
SWIR: 30 µm x 30 µm
Table 11: APEX main specifications
The APEX Science Center is located at the University of Zurich while the APEX Operation
Center is located at VITO. Flight campaign planning is coordinated from VITO. The processing
of the APEX data (geometric and atmospheric correction) is performed at VITO’s Central Data
Processing Center (CDPC) (Biesemans et al., 2007 [LIT23]).
The.in the CHB (Calibration Home Base) well-calibrated hyperspectral airborne APEX sensor
can serve as an excellent instrument to carry out an in-flight spectral calibration of PROBA-V.
The main advantages of using APEX are (Nieke, 2001 [LIT24])
1. the measured radiance of APEX and PROBA-V can be compared directly when both view
the same ground pixel at the same time (cfr. Sampling problem of field spectroradiometers)
2. the uncertainties of the atmosphere can be minimized by flying well above the boundary layer
of the atmosphere
3. in contrast to PROBA-V, APEX can be re-calibrated on the ground in the Calibration Home
Base (based at DLR)
4. no calibration panel is required
5. APEX allows fast sampling over a large calibration site and thus can be used for calibration
of low, medium as well as high resolution spaceborne sensors
This in-flight calibration experiment should be performed under stable atmospheric conditions
(cloud free, small aerosol loading) above at least one of the prime large spatial uniform earth
calibration targets (e.g. Tuz Gölü, Turkey). In case a suspicion of non-linearity of PROBA-V
detectors exists, more calibration sites should be selected covering the full dynamic range (water,
sand, ice,…). The combination of both lower-reflectance and higher-reflectance targets improves
the quality of the calibration (Teillet et al., 2001 [LIT25]).
In the direct method, the radiance values measured by APEX at airborne level (e.g. at 7 km) in
the wavelength range from 380 nm to 2500 nm are converted to TOA radiances using a radiative
transfer code to take into account residual scattering and absorption between aircraft and satellite.
The APEX-based TOA radiances are then spectrally resampled to the spectral response curves of
PROBA-V and compared to the PROBA-V TOA measurements to check or compute the absolute
calibration coefficients of PROBA-V.
In order to have the same illumination conditions, APEX acquisitions should be timed to coincide
with the PROBA-V overpass. That means at Tuz Gölü at TBD (dependent on PROBA-V launch)
LST (Local Solar Time). Furthermore, to have the same viewing conditions for PROBA-V and
APEX, the nadir centre lines should coincide as well. In this configuration only the centre
PROBA-V sensor can be calibrated. To be able to calibrate also the left and right sensor, the
PROBA-V sensor will be tilted (roll manoeuvre) +17,5° which will allow nadir viewing of the
left and centre sensor (overlapping pixels) and subsequently in a next PROBA-V overpass tilted
(roll manoeuvre) -17,5° which will allow nadir viewing of the centre and right sensor
(overlapping pixels).
At 7 km altitude above ground level (AGL), the APEX FOV (+/- 14°) results in a swath width of
3491 m and 3.5 m pixels. A swath width of 3491 m corresponds to approximately 34 100 m
PROBA-V pixels.
Averaging of pixels (34x34 PROBA-V pixels for the BLUE, RED and NIR and 11x11 PROBA-
V pixels for the SWIR) is required to reduce uncertainties due to inhomogeneities of the
calibration site and mis-registration between APEX and PROBA-V.
Due to high operational costs, calibration on the basis of aircraft underflights cannot be
considered as a fully operational method for the exploitation phase but will be performed as
validation once or twice a year.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 55/139
4.4.3.2 Algorithm
4.4.3.2.1 Physics of the algorithms
There are two methods to determine TOA radiance from APEX data. The indirect method and the
direct method:
The direct method compares directly the TOA PROBA-V radiance to the APEX radiance
corrected for the remaining atmospheric contribution between APEX and PROBA-V without
using additional ground-truth data.
The indirect comparison method starts from APEX radiance (the sun-reference site-APEX
geometry should be known) which is converted to surface reflectance using a radiative transfer
model (MODTRAN, 6SV) taking into account the viewing and observation geometry and
atmospheric properties.
The direct method without using ground-truth instruments may not reach the same accuracy as
the indirect method (Hovis et al., 1985 [LIT26]). Therefore we propose to follow the indirect
comparison method of which the physics is described below:
First the APEX scan lines of a segment l are averaged according:
J
j
vsl
reflji
k
vsllirefk L
JL
1
,,,, ),,(
1),,(
Equation 24
with v view zenith angle, s solar zenith angle, relative azimuth angle between solar azimuth
and view azimuth direction, i image pixel, j image line, k spectral band index, l flight line
segment and J total number of flight lines in an image segment.
Ground truth measurements of aerosol optical thickness and water vapour, reflectance and BRDF
are taken to constrain the atmospheric correction and to characterize the calibration site.
For the atmospheric characterization the Langley method as already described in 4.4.2.2.2.1 is
used to determine aerosol optical depth, Angström coefficient and water vapour content.
The average radiance measured by APEX ( l,iref,kL ) is then atmospherically corrected to obtain
the average surface reflectance. This is done at VITO’s Central Date Processing Center (CDPC)
for airborne hyperspectral data (Biesemans et al., 2007 [LIT23]). The input parameters for the
MODTRAN-based atmospheric correction performed in the CDPC are visibility or aerosol
optical depth (AOT550), ozone and water vapour concentration, terrain elevation, aircraft altitude
and sun-viewing geometry. After atmospheric correction (f-1
) the surface reflectance averaged
over the scan lines is obtained.
J
j
vsl
reflji
k
vslli
refk Lf1
,,1
,
, ),,(),,(
Note that the surface reflectance is still a function of bidirectional reflectance of the surface.
In a next step the surface reflectance is adjusted to nadir view angle and an average solar zenith
angle during APEX data acquisition.
BRDFvsl
refk
srefk
vslli
refksli
refk
),,(
)0,0,(),,()0,0,(
,
,
,
,
,
,
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 56/139
Equation 25
Bidirectional reflectance factor adjustment can be determined from goniometer measurements of
the calibration site.
The BRDF corrected pixels are then averaged to obtain a single reflectance for one flight line
segment with I the total number of pixels in a scan line
I
i
s
refli
k
sl
refk
I 1
,, )0,0,(
1)0,0,(
Equation 26
In case more than one flight line segment cover the calibration site, one has to average over the
total number of flight line segments L to obtain a single averaged reflectance for the entire
calibration site.
L
l
s
refl
k
s
refk
L 1
, )0,0,(1
)0,0,(
Equation 27
Starting from the single averaged calibration site reflectance (standardized for nadir viewing and
average solar zenith angle), one can estimate the TOA radiance that PROBA-V should see
following three steps: i) BRDF adjustment of the standardized surface reflectance for the
PROBA-V geometry; ii) conversion of reflectance to TOA radiance by adding atmosphere
(inverse atmospheric correction; iii) convolution with PROBA-V spectral response function.
The BRDF adjustment of the standardized APEX-reference surface reflectance for the PROBA-V
geometry (sat refers to the solar and viewing angles pertinent to PROBA-V imaging the
calibration site) can be written as:
BRDFs
refk
satsatvsats
refk
s
refksatsatvsats
refk
)0,0,(
),,()0,0,(),,(
,
,,
,
,
,,
,
Equation 28
The BRDF adjustment factor need to be determined for the different PROBA-V spectral bands.
The conversion of reflectance to TOA radiance using a Radiative Transfer Model is represented
by
),,(),,( ,,
,
,,
,satsatvsats
refksatsatvsats
refk fL
Equation 29
with the Radiative Transfer Model denoted with f for simplicity.
The same atmospheric parameters as for the APEX atmospheric correction are used while
allowing a sensor altitude difference. (BRDF is taken into account in 6SV). This two-way use of
the same aerosol optical depth for the APEX-based surface reflectance retrieval and the TOA
radiance prediction results in a low sensitivity to aerosol optical depth (0.5% increase in TOA
radiance for an aerosol optical depth which is double the amount on the APEX flight (Teillet et
al., 2001 [LIT25]).
The final step is the convolution of the predicted TOA radiance with the PROBA-V spectral
response function as follows:
0
,,
, /(),,()( dSdSLBLUEL BLUEBLUEsatsatvsats
refkref
TOA
0
,,
, /(),,()( dSdSLREDL REDREDsatsatvsats
refkref
TOA
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 57/139
0
,,
, /(),,()( dSdSLNIRL NIRNIRsatsatvsats
refkref
TOA
0
,,
, /(),,()( dSdSLSWIRL SWIRSWIRsatsatvsats
refkref
TOA
Equation 30
The change in calibration coefficients for the BLUE, RED, NIR (SWIR) bands is calculated
based on a comparison of the measured PROBA-V radiance averaged over all 34x34 (11x11)
PROBA-V pixels and APEX-based TOA radiance as:
)(
)(
BLUEL
BLUEL
A
AA
ref
TOA
meas
TOA
old
BLUE
new
BLUEBLUE ;)RED(
refTOA
L
)RED(measTOA
L
oldSWIR
A
newSWIR
ASWIRAΛ
)NIR(refTOA
L
)NIR(measTOA
L
oldSWIR
A
newSWIR
ASWIRAΛ
)SWIR(refTOA
L
)SWIR(measTOA
L
oldSWIR
A
newSWIR
ASWIRAΛ
Equation 31
4.4.3.2.2 Algorithm implementation
4.4.3.2.2.1 Processing
APEX processing
Extract reference site (segment) from georeferenced APEX image
Clear pixel selection: pixels affected by clouds, cloud shadow, haze are removed in the CDPC
Average scan angle image: the scan lines of the segment are averaged
Atmospheric correction with MODTRAN (CDPC) using AOT estimate, Angström
coefficient and water vapour content from sun photometer measurements to retrieve surface
spectral reflectance as a function of view angle
BRDF correction: to become reflectance for nadir view angle and average solar zenith angle
Average surface reflectance for all pixels in scan line
Repeat previous steps in case more than one APEX flight line is needed to cover the reference
site
BRDF correction of standardized surface reflectance for averaged PROBA-V geometry
Simulate atmosphere: Conversion with MODTRAN of BRDF corrected surface reflectance
to TOA radiance by adding atmosphere using same AOT estimate, Angström coefficient and
water vapour content from sun photometer measurements as used for the APEX atmospheric
correction and using the same solar irradiance.
Convolution with PROBA-V spectral response function
PROBA-V processing
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 58/139
Extract calibration site pixels
Conversion of DN to meas
TOAL : The input DNs for all selected image pixels are converted to
TOA radiance meask
TOAL ,following Equation 4.
Clear pixel selection : Visual inspection for clouds. The APEX-based method is only applied
under optimal weather conditions (cloud-free).
Averaging PROBA-V radiance for near nadir calibration site pixels (34x34 pixels for BLUE,
RED, NIR, 11x11 pixels for SWIR)
Comparison of averaged APEX-based TOA radiance with averaged nadir PROBA-V TOA
radiance
4.4.3.2.2.2 Required ancillary data
Goniometer for BRDF measurements to determine BRDF correction factor for reference site
(available from CEOS by 2012)
Sun photometer to determine Aorosol Optical Thickness, Angström coefficient and water vapour
content
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 59/139
4.4.3.2.2.3 Summary of the processing steps
Figure 10: Flowchart to calculate calibration coefficients
NL(DNjk ) - DCjm
k
LTOAjk,meas
= --------------------------------------------------
Ak * Gm
k * gjm
k
DNk Current calibration file (DN
k
Ak ,G
k,gi
k), date,θs,E
k
loop on line j
loop on pixel
i
end of loop on
pixel i
end of loop on line
j
LTOAjk,meas
Cloud (visual inspection)
No
Averaging (TBDxTBD pixels for BLUE, RED, NIR, SWIR)
LTOAjk,meas
Discard pixel
Yes
ABLUE
ARED
ANIR
ASWIR
Ground reference
reflectance
Rho
water vapor, AOT550,
Angstrom coefficient
Radiative Transfer
Model
Convolution
PROBA-V SRF
Simulated
LTOAk,ref
(based
on ground
reflectance)
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 60/139
Figure 11: Flowchart for APEX-based TOA radiance for PROBA-V spectral bands
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 61/139
4.4.3.3 Error analysis
The uncertainty has been estimated at 2.8% for the radiance-based indirect method (Biggar et al.,
1994 [LIT21]). A precision of 1.8% can be obtained with new instrumentation (Dinguirard and
Slater, 1999 [LIT20]). The major contribution to uncertainty is related to the calibration of the
airborne spectroradiometer. The error given is the error in the TOA radiance caused by the error
source. All percentages are 1 . We assume the error sources are independent such that the total
error is the root sum of squares of the errors caused by all listed error sources.
Error Sources Error
Radiometer calibration 2.5
Measurement accuracy 1.3
Correction for altitude difference <0.1
Total error (root sum of squares) 2.8
Table 12: Error sources for radiance-based method with reference to NIST standards.
4.4.4 Absolute calibration over stable deserts
4.4.4.1 Desert Sites selection
Deserts are well suited as test sites because they are usually very stable over time (good for multi-
temporal calibration) and are seldom covered by clouds. Also they are spatially homogeneous.We
compile a list of criteria based on Cosnefroy et al., 1996 [LIT38] and also note their requirements
for suitable test sites next to them in . In line with these criteria, a set of 20 suitable test sites have
been selected in North Africa and Saudi Arabia. They are shown on a map in (reproduced from
(Cosnefroy et al., 1996 [LIT38]) and listed in .
Criterium Requirement
Spatial homogeneity 2 or 3% dispersion threshold on TOA
reflectance over 100x100 km area
Temporal stability 15- 20 % or better at seasonal scale, after
correction for atmospheric effects
Low directional effects < 15 %, based on AVHRR data
Low cloud coverage > 50 % clear days annual
Low precipitation < 10 mm /month
Table 13: Selection criteria and requirements for suitable desert sites
.1 Country & number .2 Latitude .3 Longitude
.4 Algeria 1 .5 23,80 .6 -0,40
.7 Algeria 2 .8 26,09 .9 -1,38
.10 Algeria 3 .11 30,32 .12 7,66
.13 Algeria 4 .14 30,04 .15 5,59
.16 Algeria 5 .17 31,02 .18 2,23
.19 Mauritania 1 .20 19,40 .21 -9,30
.22 Mauritania 2 .23 20,85 .24 -8,78
.25 Mali 1 .26 19,12 .27 4,85
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 62/139
.1 Country & number .2 Latitude .3 Longitude
.28 Niger 1 .29 19,67 .30 9,81
.31 Niger 2 .32 21,37 .33 10,59
.34 Niger 3 .35 21,57 .36 7,96
.37 Libya 1 .38 24,42 .39 13,35
.40 Libya 2 .41 25,05 .42 20,48
.43 Libya 3 .44 23,15 .45 23,10
.46 Libya 4 .47 28,55 .48 23,39
.49 Egypt 1 .50 27,12 .51 26,10
.52 Sudan 1 .53 21,74 .54 28,22
.55 Arabia 1 .56 18,88 .57 46,76
.58 Arabia 2 .59 20,13 .60 50,96
.61 Arabia 3 .62 28,92 .63 43,73
Table 14: Desert sites used for radiometric calibration
Figure 12: Location of desert sites used for radiometric calibration,
The test sites are all covered by the normal operational imaging of PROBA-V. This has the
advantage that they are imaged almost daily. No special settings are be used for their observation.
Only clear pixels can be used for the analysis. If a substantial portion of the site is cloud covered,
the observations for that site should not be used altogether for that day.
4.4.4.2 Desert Calibration Methodology
Desert calibration relies on the comparison between TOA reflectance as measured by the
PROBA-V sensors meas
TOA and modelled TOA reflectances values el
TOA
mod for these targets.
4.4.4.2.1 Calculation of modelled TOA reflectance values
Reference or modelled TOA reflectactance values (el
TOA
mod ) will be calculated using the radiative
transfer code 6SV.
The modelled TOA reflectance values are determined by three types of factors:
o surface BRDF properties
o atmospheric conditions
o observation conditions including solar zenith angle, view zenith angle and
relative azimuth angle.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 63/139
4.4.4.2.1.1 Surface properties
The surface properties of desert sites can be described well with a simple BRDF model. We
assume the surface properties to be stable over time, so they do not have to be updated. From these
TOA radiances can be calculated for any given illumination and viewing condition.
The bi-directional reflectance (BRDF) of the surface of the desert sites can be modelled using a
reflectance model. We use the standard Rahman-Pinty-Verstraete (RVP) model from Rahman et
al., 1993 [LIT39]. It describes the reflectance in function of the direction of the illumination (ss)
and viewing (vv), using only three parameters (0 , k and ):
)(1)(.)cos(cos
coscos),;,(
1
11
0 gRgFk
vs
v
k
s
k
vvsss
.
0 and k are two empirical surface parameters. 0 characterizes the intensity of the reflectance of
the surface, whereas k describes the anisotropy of the surface. F is the function:
2/322
2
)cos(21
1)(
ggF
in which controls the forward and backward scattering. The phase angle g is given by
)cos(sinsincoscoscos vsvsvsg
The hot spot effect is approximated by:
GGR
1
11)(1 0
with
)cos(tantan2tantan 22
vsvsvsG
Reference reflectances should be available for every test site for a wide range of viewing angles.
We make use of a descriptive database which accurately describes the 18 of the 20 test deserts
sites (compiled by Govaerts Y., Eumetsat; not including Mauritania 1 and Arabia 3). For every
desert, the database contains for all available angles the information:
1. wavelength
2. 0 parameter
3. asymmetry parameter
4. k parameter
5. hotspot parameter
4.4.4.2.1.2 Atmospheric properties
The main atmospheric parameter affecting the TOA reflectance are :
1. Atmopheric pressure
2. Ozone
3. Water vapour
4. Aerosol model and aerosol concentration
The first 3 will come from ECMWF meteo data.
The 6SV desert aerosol model is currently used. The AOT variability over the desert sites is
mainly seasonal as can be seen in the few Aeronet sumphotometer stations nearest to the deserts
sites (Figure 20). During ‘winter’ months (October to February) AOT values for all Aeronet site
are significantly lower than the summer AOT values, similar to the monthly variation as used in
Govaerts and Clerici 2004, [LIT1] (Cyan coloured line in figure). The team at CNES is however
using in their calibration over deserts a fixed AOT value of 0.2 (Red coloured line in figure).
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 64/139
Both options i.e. a montly variable AOT value following Govaerts and Clerici 2004 and a fixed
AOT value following CNES are currently being evaluated (Up to now prototyping results have
indicated no major difference, but aerosol issues will be further investigated in collaboration with
Y. Govaerts).
Figure 13: AOT seasonal variation
4.4.4.2.2 Calculation of measured TOA reflectance values
Conversion to DNs to TOA radiance meas
TOAL by applying current calibration coefficients
following Equation 4.
Conversion of TOA radiance meas
TOAL to apparent TOA reflectance meas
TOA (Equation 8)
4.4.4.2.3 Cloud check
Conversion to DNs to TOA radiance meas
TOAL by applying current calibration coefficients
following Equation 4. The cloud masking is based on a threshold test, if one pixel in the ROI is
cloudy, the whole ROI is discarded.
if )(*)( ,,,, bluemeas
TOA
NIRmeas
TOA
bluemeas
TOA
NIRmeas
TOA threshold than pixel = cloudy.
We only proceed with desert sites that are completely cloud-free. This is less complex and is
justified as the desert sites are expected to be cloud-free for a sufficient large portion of the
time.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 65/139
4.4.4.2.4 Averaging over the ROI
All TOA reflectance values for all pixels in the ROI are averaged to get one single value for the
ROI for each image. This is justified as the ROI is relatively small and homogeneous.
4.4.4.2.5 Check on VZA
As the uncertainty of the BRDF model increases with VZA, scenes obtained with a VZA larger
than 30° are removed.
4.4.4.2.6 Comparison of TOA reflectance values
The TOA reflectance values as measured by the sensor meas
TOA can now be compared to the
modelled values el
TOA
mod for e ach desert site. A new estimate of the calibration coefficient for
that site and day is then calculated as :
el
TOA
meas
TOAoldnew AAmod
4.4.4.2.7 Outlier selection and daily averaging
A daily average is calculated by averaging the results obtained over the different sites after outlier
removal. A site is removed if it is detected as an outlier in at least one of the spectral bands. A
robust outlier selection procedure based on the median and standard deviation from the median is
used for this (see also 4.6.2.2).
4.4.4.3 Error sources
The determination of the calibration coefficient for individual ROIs pixels is affected by a
number of variables, each of which is prone to random and/or systematic errors. The
contributions to the errors come mainly from the following sources, which we will describe
briefly:
Properties of the target: The BRDF surface properties of the desert sites are assumed to be
constant over time. Even when maximally stable targets are selected, this assumption is only
approximately true, therefore it is expected to contribute significantly to the errors.
Accuracy of the BRDF model: The BRDF model derived by Govaerts and Clerici 2004, [LIT1]
dates from 2004 and has been compiled on the basis of inversion of satellite data (which have
their own uncertainties) and other datasources available at that moment. The BRDF dataset
contains therefore some intrinsic uncertainties. A BRDF data set with improved accuracy
compiled on the basis of MODIS data may become available in the near future (personal
communction Yves Govaerts).
Aerosol model: The standard 6SV desert aerosol is currently used. This model (i.e. its scattering,
absorption ,asymmetry parameters) may deviate from the actual aerosol.
Aerosol optical thickness: Aerosol optical measurements in these deserts sites are rare and
therefore assumptions have been made on the used AOT (i.e. monthly variable or fixed AOT)
which may deviate from the actual aerosol optical thickness. It is expected that after averaging a
lot of measurements over sites and over time, that the aerosol affect will be more a random error
(noise) than a fixed bias/systematic error.
Radiative transfer code: The intrinsic uncertainty of the 6SV radiative transfer code may cause
systematic uncertainties in the absolute calibration results.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 66/139
Seasonal variation : The 4th degree seasonal polynomial correction function is derived on the
basis of SPOT-VGT data analysis (only after at least one year of PROBA-V acquisitions, this can
be replace by a PROBA-V derived function). Hereby it is assumed that the observed seasonal
variation in the SPOT-VGT data analysis is only due to deserts related effects and not to the
sensor itself. Furthermore, it is assumed that the seasonal variation pattern is similar over the
years. Possible small violation of both assumptions will give rise to errors in the absolute
calibration on the basis of stable deserts.
Uncertainties in ECMWF meteo information: About 5-10 mbars is a reasonable accuracy for
the atmospheric pressure. An accuracy of about 10-20 mAtm-cm (Dobson Units, DU) is a typical
accuracy for the total ozone amount of the atmosphere. An uncertainty of 20 % is often used for
the water vapour content.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 67/139
4.5 Relative radiometric calibration
4.5.1 Interband: Sun glint
4.5.1.1 Introduction
This method uses the specular reflection of the sun on the ocean surface. This sun glint reflection
is high and spectrally flat and is used to transfer the absolute calibration of one reference band to
other spectral bands (inter-band calibration) (Hagolle et al., 2004 [LIT27]; Fougnie et al, 2007
[LIT4]). The size of sun glint spot is variable as it depends on the ocean surface roughness which
is controlled by wind speed. The range of angles from which the sun glint can be observed is
larger for an agitated sea. The sun glint spot observed by a satellite at an altitude of 800 km often
exceeds 100 km. The main advantage of the sun glint calibration approach is that it is one of the
rare methods that can provide calibration of the SWIR band. Due to the chosen local time (around
10h30) of the descending node in the sun-synchronous orbit of PROBA-V, sun glint is always
observed in eastern direction. It can only be observed by the eastern and middle looking cameras
in normal operating mode (depending on location and day of the year) because their viewing
direction is close to the exact specular direction. Observation of sunglint by the western looking
camera is not possible without platform manoeuvres.
4.5.1.2 Algorithm
4.5.1.2.1 Physics of the algorithm
4.5.1.2.1.1 Conversion to TOA reflectance
See section 4.4.1.2.1.1
4.5.1.2.1.2 TOA reflectance signal decomposition
The TOA reflectance cTOAρ corrected for gaseous transmittance, over oceans in specular (sun
glint) conditions can be decomposed as (omitting spectral band and angles for simplicity):
w
vtotstotw
vtotstotspe
gas
TOAc
TOAs
TrTrTrTr
Tr path
1
,,
,,
Equation 32: Sun glint TOA reflectance
with
speρ : the specular reflectance at the ocean surface, for the other terms see 4.4.1.2.1.2.
We ignore here the surface foam reflectance, which is negligible for surface wind speeds under
10 m/s (wind speed mask applied in processing). Compared to equation Equation 9 of section
4.4.1.2.1.2 Equation 32 includes explicitly the sun glint contribution.
For a flat sea surface (zero wind speed) the specular reflectance speρ or directly reflected light
can be computed ‘exactly’ using the Snell-Fresnel laws. For a rough sea surface, the reflection is
conditioned by the wind and therefore the sun glint reflectance of the sea surface can only be
described on statistical basis. Cox and Munk [LIT28] took many photographs of sun glitter
patterns on the ocean surface under different conditions of wind speed/direction and sun angles.
From these observations they derived an ocean wave slope probability distribution function
(PDF) that can be described as a Gaussian curve plus higher-order skewness and kurtosis terms.
A first approximation of the Cox-Munk model uses an isotropic Gaussian slope distribution
(isotropic rough surface, independent of wind-direction) to represent the oceanic wave slopes.
Adopting this isotropic form, the probability of a spatial sample being contaminated by sun glint
is given by :
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 68/139
²coscos²
²coscos)cossinsincoscos1(2exp(
²
1
sv
svvsv
sg
sP
where sθ is the sun zenith angle (SZA) at the viewed spatial sample
vθ is the view zenith angle (VZA) at the viewed spatial sample
φΛ is the relative azimuth angle (RAA) (sun azimuth angle (SAA)-view azimuth
angle (VAA)) at the viewed spatial sample
²σ is the mean surface slope which is function of the wind speed ws
wsσ 00512.0003.0²
This isotropic wave slope PDF is often used in remote sensing applications when wind direction
is not accurately known or not uniform.
In 6SV the slope distribution is considered anisotropic: the distribution of the slope components
depends on the wind direction. The wave slope PDF in 6SV is expressed with Gram-Charlier
expressions as :
2exp
2
,),(
224''
uc
yxsg
GZZP
with
and the normalized x and y slopes defined as
u
xZ
'
, c
yZ
'
c and u are the facet slope standard deviations in the crosswind and upwind :
004.000192.0003.02 wsc , 002.000316.02 wsu
,4G the Gram-Charlier expansion :
36
11
2
11, 3
03
2
214 ccG
3624
111
4
136
24
1 24
04
22
22
24
40 ccc
with 21c and 03c the skewness coefficients defined as
03.00086.0001.021 wsc and 004.00033.0004.003 wsc
and the kurtosis coefficients (04,40,22 ccc ) are
23.040.040 c , 12.012.022 c , 41.023.004 c
and
'xZ and '
yZ are functions of xZ and yZ the crosswind and upwind directions of the slope (see
Figure 14) given by
tansinxZ , tancosyZ
is facet azimuth angle (clockwise from the sun) and is the facet zenith angle (tilt). Let χ be
the wind direction in the local frame (related clockwise from the North by w , then
ws
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 69/139
) and ws the wind speed in m/s (at a height of 10 m). If the sun system (x,y) is rotated through an
angle χ to a new system (x’,y’) related to the wind direction, then the facet slopes 'xZ and '
yZ in
this wind system are
tan'sin' xZ , tan'cos' yZ , '
therefore
yxx ZZZ sincos', xyy ZZZ sincos'
Figure 14: Sun glint geometry
In Figure 14 sθ is the sun zenith angle, vθ is the view zenith angle, φΛ is the relative azimuth
angle, is facet azimuth angle, is the facet zenith angle, W wind speed and ws .
The uncertainties in the Cox and Munk parameters are quite large. For a windspeed of 10 m/s, the
relative error or uncertainty in 2cσ and 2
uσ is about 10% and 50 % for 21c , 03c , 40,22 cc . The most
uncertain value is 04
c which has an error of 200 %.
In order not to considerably complicate the sun glint calibration approach by including wind
direction and because accurate information of wind direction just above the sea surface is often
unavailable for the time of observation, it is decided to modify 6SV slightly by either taking the
average speρ obtained by looping over the wind azimuth from 0° and 360° or by replacing the
PDF in 6SV by the isotropic formulation.
The sun glint (specular) reflectance is given by:
sg
vs
ssvs
spe Pnr
4coscoscos4
,,,,
Equation 33: Sun glint reflectance
where ssvsnr ,,,, is the Fresnel reflection coefficient (with n complex index of
refraction of sea water)
is the angle formed by the reflecting facet normal and the local normal defined
by
2cos22
coscoscos
sv
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 70/139
where is the specular reflection angle defined by
cossinsincoscos2cos svsv
The Fresnel reflection coefficient (r) describes the proportion of light hitting the surface that is
reflected back. It can be calculated taking into account the refraction indices of sea water and air
and the angles of incidence and refraction. In 6SV the fresnel reflectance is computed according
to (Born & Wolf, 1975 [LIT29] taking into account both the real and imagery index of refraction.
The refractive index of water varies with wavelength. In 6SV the complex index of refraction of
sea water is deduced from the complex index of refraction of pure water, specified by Hale &
Querry (1973) [LIT30].An additional correction of +0.006 is added to the real component due to
the salinity and chlorinity of typical seawater. The small variation of the refraction index in
function of the sea water temperature is neglected. In Table 15 the real and imagery part of the
index of refraction are given for the PROBA-V wavelengths. Uncertainties in the index
refraction are taken into account in the error budget (see section 4.5.1.3).
Band nr (real component index of
refraction)
nr (imagery component index of
refraction)
BLUE 1.342 1E-09
RED 1.337 2.03E-08
NIR 1.335 2.68E-07
SWIR 1.323 8.55E-05
Table 15: Sea water index of refraction for the PROBA-V bands
4.5.1.2.1.3 Retrieval of the wind speed from the RED band
Because of the uncertainty on wind from ECMWF global models (typically 2 m/s on wind
speed) and the error bars of the Cox and Munk model, there is an uncertainty in the
determination of speρ . Therefore the sun glint calibration approach can’t be used with enough
accuracy for absolute calibration of all bands. However it is adequate for inter-band calibration,
using a well-calibrated band as reference band to derive the wind speed from the image pixel
itself by comparing image data with simulated data.
The reference is preferably chosen between the bands that are calibrated with the Rayleigh
calibration method, i.e. the BLUE and RED band. The RED band is the preferred reference band.
In the BLUE c
TOA varies less with wind speed as there the relative contribution of speρ to c
TOA is
lower. Therefore wind speed retrieval from the data itself will work best with the RED band.
More specifically, the wind speed will be derived based on a LUT of )(REDc
TOA versus wind
speed generated with 6SV (see section 4.5.1.2.2.1). The wind speed value for which the
measured )(, REDmeasc
TOA agrees best with the modelled )(mod, REDelc
TOA is selected
As described in the technical note on Sun Glint [PVDOC-611], due to the difference in viewing
geometry between the different bands, especially between VNIR and SWIR detectors, the
uncertainty in the Cox-Munk model is not fully cancelled for the case of PROBA-V. The
uncertainty in the ratio of sun glint reflectances for two different relative azimuth angles
SWIRRAA
spρREDRAAsp
ρ introduces still an error in the interband sun glint calibration
approach (see section 4.5.1.3)..
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 71/139
4.5.1.2.1.4 Retrieval of the calibration coefficients for the BLUE, NIR and SWIR bands
Once the wind speed has been determined from the observations in the RED band, this wind
speed value is then used to model, with 6SV,
)(mod, BLUEelc
TOA , )(mod, NIRelc
TOA , )(mod, SWIRelc
TOA .
The change in calibration coefficients for the Blue, NIR, SWIR is then calculated as :
)(
)(mod,
,
BLUE
BLUE
A
AA
elc
TOA
measc
TOA
old
BLUE
new
BLUEBLUE
)(
)(mod,
,
NIR
NIR
A
AA
elc
TOA
measc
TOA
old
NIR
new
NIRNIR
;
)(
)(mod,
,
SWIR
SWIR
A
AA
elc
TOA
measc
TOA
old
SWIR
new
SWIRSWIR
4.5.1.2.2 Algorithm implementation
4.5.1.2.2.1 Generation of LUTs
For the generation of the LUTs 6SV is used. 6SV takes into account coupling effect between
ocean and atmosphere (including Cox Munk model), polarization and multiple-scatttering which
are necessary to accurately calculate elc
TOA
mod, . To calculate these LUTs of elc
TOA
mod, the following
parameter settings are used:
- the aerosol model and aerosol optical thickness are fixed. The Shettle and Fenn Maritime
aerosol with 98% humidity (denoted as M98, see section 4.4.1.2.1.3) is used with a fixed
AOT of 0.08 at 850 nm. The spectral dependency of AOT ( αλ )for the M98 aerosol,
expressed by the angstrom coefficient a, is 0.1. This results in an AOT of approximately
0.0836 at 550 nm (which is the wavelength to express AOT in 6SV). Sensitivity studies by
Hagolle et al (2004) [LIT27] have shown that for this aerosol model the calibration method is
quite tolerant to a bad knowledge of AOT. To discard observation with too high aerosol
loading or an aerosol type very different from M98, the MODIS Terra Aerosol Products (or
equivalent sensor if MODIS is not available) will be checked, containing both AOT and
aerosol size distribution. The local equatorial crossing time of the Terra satellite is
approximately 10:30, which is very close to the PROBA-V overpass.
- The water leaving reflectance wρ will be modelled in 6SV in function of chlorophyll
concentration according to the Morel bio-optical model. As the same oligotrophic sites (see
Table 6) as used for the Rayleigh calibration are selected for the sun glint calibration, also the
same monthly mean chlorophyll concentration will be used. Same sites are used to minimize
the uncertainty in chlorophyll concentration. Furthermore by taking monthly average values
errors due to seasonal chlorophyll variations are minimized.
For the different PROBA-V bands LUTS are created for the apparent TOA reflectance,
corrected for gaseous transmittance for an atmosphere bounded by fresnel reflecting wind
roughened oceanic surface with a chlorophyll concentration taken from
. These LUTs are a function of sun zenith angle, view zenith angle, relative azimuth angle and
windspeed.
Two options for the generation of the LUTs with 6SV are considered: (a) a global pre-calculated
LUT and (b) an on-the-fly calculated LUT. The global LUT will be valid for all imagery and
contains cTOAρ for a variety of sun-view geometries. The on-the-fly calculated LUT is much
smaller and only valid for the image under consideration.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 72/139
4.5.1.2.2.2 Processing steps
Conversion to DNs to TOA radiance meas
TOAL by applying current calibration coefficients
following Equation 4.
Conversion of TOA radiance meas
TOAL to apparent TOA reflectance meas
TOA (Equation 8)
Correct for the gaseous absorption (conversion of meas
TOA to measc
TOA
, ) based on predefined
exponential variation of gaseous transmittance with airmass and gaseous amount for each band
(SMAC approach).
Visual cloud screening: A visual cloud inspection is considered to give higher performance
than an automatic cloud detection. It the sun glint spot is cloudy, the sun glint scene not be
used
White caps masking: Remove all pixels, probably contaminated by surface white caps based
on wind speed threshold (5 m/s) (derived from the nearest meteo data)
Sun glint test: Automatically select sunglint spot based on a wave angle threshold θn. All
pixels for which θn < 4° are kept with θn defined as :
2cos2
coscoscos
p
vs
n ar
with cossinsincoscoscos vsvsp ar
Homogeneity test : remove pixels with large local variation as this may be due to the presence
of some clouds not detected by the visual cloud screening algorithm : remove pixels for which
stdev( )(*1.0)55( ,, NIRc
TOA
NIRc
TOA avgpixelsx . An extra test based on the raito of NIRc
TOA
SWIRc
TOA
,
,
used for SPOT-VGT to indicate the presence of undetected clouds is not applicable for
PROBA-V as due to the view angle differences between VNIR and SWIR the specular
reflectance ratio is signifcantly different from one even for uncloudy conditions.
NIR reflectance test: remove all pixels for which 2.0, NIRc
TOA to select only those pixels for
which the sun glint reflectance is high enough to minimize perturbations linked to ocean
surface or atmospheric effects;
Wind speed estimation : Use global LUT of kc
TOA
, or on-the-fly calculated LUT of kc
TOA
, in
function of wind speed VZA, SZA,RAA for the chlorophyll concentration taken from
and parameter ranges valid for the selected pixels. For all LUTS AOT is fixed at 0.08 and
M98 aerosl is used. For each pixel find windspeed (wsest
): the simulated )(mod, REDelc
TOA (in the
global LUT or the on-the-fly calculated LUT) are interpolated on the angles from each pixel to
obtain )(mod, REDelc
TOA corresponding to the geometry of the observation. Then
),(mod,
iwsREDelc
TOAρ and ),( 1
mod,
i
elc
TOA wsRED that surround )(, REDmeasc
TOA are found and
assuming a linear relationship the wind speed (ws) corresponding to )(, REDmeasc
TOA is
estimated.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 73/139
For each pixel find ),,(mod,
SWIRNIRBlueelc
TOAρ corresponding to ws
est :
),,,(mod,
wsSWIRNIRBlueelc
TOAρ is derived from a linear interpolation between
),,,(mod,
iwsSWIRNIRBlueelc
TOAρ and ),,,( 1
mod,
i
elc
TOA wsSWIRNIRBlue at the estimated wind
speed wsest
Calculate for each pixel the change in calibration coefficients for the Blue, NIR, SWIR as :
)(
)(mod,
,
BLUE
BLUE
A
AA
elc
TOA
measc
TOA
old
BLUE
new
BLUEBLUE
;
)(
)(mod,
,
NIR
NIR
A
AA
elc
TOA
measc
TOA
old
NIR
new
NIRNIR
;
)(
)(mod,
,
SWIR
SWIR
A
AA
elc
TOA
measc
TOA
old
SWIR
new
SWIRSWIR
Discard scenes for which MODIS Terra (if available after 2012) AOT is higher than 0.15 and
angstrom exponent is higher than 0.3
4.5.1.2.2.3 Required ancillary data
Some exogenous data are required to accurately calculate elc
TOA
mod, , to correct meas
TOA for gaseous
transmittance or in the masking processing steps. These are :
Wind speed at sea level to mask white caps: this parameter can be obtained from ECMWF
with an accuracy of about 2 m/s
Total ozone amount to calculate the ozone absorption: this parameter can be retrieved from
ECMWF or TOMS with an accuracy of 10-20 mAtm-cm
The water vapour content obtained from ECMWF
MODIS Terra Aerosol product (or equivalent if not available e.g. MERIS)
Other ‘internal’ ancillary data (assumed attached to the input image) : time, date, sun zenith
angles, solar zenith angles, relative azimuth angles, cloud mask (calculated from the image)
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 74/139
4.5.1.2.2.4 Summary of the processing steps
Figure 15: Overview Flowchart DCC
4.5.1.3 Error analysis
In Hagolle et al. (2004) [LIT27] a detailed error analysis has been performed for absolute
calibration of SPOT-Vegetation using the sun glint calibration method.
The considered error sources are :
1) absolute calibration error in the reference RED band. When considering absolute aspects,
i.e., to propagate the absolute calibration to the other bands, the calibration error of the
reference band should be considered. The RED band is assumed to be calibrated with an
accuracy of 3 %
2) atmospheric pressure: an error of 15 hpa is considered
3) sea water refraction index: this directly effects the sun glint reflectance through Equation
33. It varies slightly with salinity and temperature. Especially for SWIR wavelengths
there still exists some uncertainty in the value of the sea water refraction index. An
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 75/139
uncertainty of 0.3% in this parameter is assumed for the SWIR band and 0.1% for the other
bands
4) chlorophyll content: as the water leaving reflectance is directly calculated from the
chlorophyll content an error in the chlorophyll content affects the reflectance in the BLUE
and RED bands. For NIR and SWIR bands the water leaving reflectance is 0 (independent
of chlorophyll) for deep oligotrophic oceans. For the BLUE end RED bands an uncertainty
in the water leaving reflectance of 20% is taken into account the error budget.
5) Aerosol model and content. The error budgets in Hagolle et al. (2004) [LIT27] is made
assuming that C70 cases and cases with high AOT are discarded
6) Gaseous transmittance: An error in the ozone amount and water vapor of respectively 5%
and 20% is used in the error budget
7) Polarization sensitivity : Errors due to polarization differences between bands were
considered by Hagolle et al. (2004) [LIT27] for SPOT-VGT, however they can be
neglected in the case of PROBA-V due to low polarization sensitivity in the different
bands.
8) Specific for PROBA-V: The relative uncertainty in the Cox-Munk model. This uncertainty
should not be taking into account for sensors with the same viewing geometry for all bands
like SPOT-VGT. It is however an error that has to be taken into account for PROBA-V
due to the difference in viewing angles, especially in azimuth angles between the VNIR
and SWIR bands. A detailed analysis can be found in the technical note on sun glint
[PVDOC-611]. A 5% relative uncertainty is considered in this study.
The linear interpolation that will be needed to intermediate between the entries of the lookup
tables should not introduce significant errors, provided that each entry of the lookup tables has
been adequately sampled (2° for θs and θv, and 5° for Δφ, 0.5 m/s for wind speed, TBC).
In Table 16 the error budget is given for the sun glint method (partly taken from Hagolle et al.
(2004) [LIT27]). Errors are 3σ values expressed in percent. The error sources are supposed to be
non corelated and thus the total error is the quadratic sum of all errors. As the spectral response
curves of PROBA-V are almost identical to those of SPOT-VGT, errors in the sun glint
calibration introduced by most of the uncertain parameters will be similar. Once the exact spectral
response curves for PROBA-V are defined, this error budget will be re-evaluated.
In section 0 it is explained how the uncertainty calculation is used in the determining the final
error budget of the vicarious calibration.
Error Sources BLUE NIR SWIR
calibration error 3% in RED band 1.7 3.2 3.3
Water reflectance 2 0.1 0.1
uncertainty sea water refraction index 0.3 0.6 2
ozone (5%) 0.3 0.5 0.3
water vapor (20 %) 0.1 0.8 0.1
Aerosols 1.5 1 2
5% relative uncertainty Cox-Munk 0 0 5
Total 3.1 3.5 6.6
Table 16: Summary of sun glint absolute calibration error budget
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 76/139
4.5.2 Interband: deep convective clouds
4.5.2.1 Introduction
This approach (Lafrance, et al., 2002 [LIT31]; Henry and Meygret, 2001 [LIT32]; Sohn et al.,
2009 [LIT33]) makes use of large, bright, thick, high altitude, convective clouds over oceanic sites,
called deep convective clouds (DCC). Their reflective properties are spectrally flat in visible and
near-infrared and the only significant contributions to the observed signal are from the cloud
reflectance, molecular scattering and ozone absorption which can be modelled with RTF. Using the
RED band as reference the BLUE and NIR band can be inter-calibrated. The method is not suited
for the SWIR band as the reflectance is no longer invariant over this spectral region.
4.5.2.2 Algorithm
4.5.2.2.1.1 Conversion to TOA reflectance
See section 4.4.1.2.1.1.
4.5.2.2.1.2 TOA reflectance signal decomposition
The TOA reflectance of deep convective clouds (DCC) over oceans can be decomposed as :
n
r
n
a
n
ccgasTOA Tr
with gasTr the total gaseous transmittance, c the reflectance at the top of the cloud/atmosphere
system and n
r
n
a
n
c ,, respectively the clouds, aerosol and rayleigh optical thickness in the
different atmospheric layers.
DCC clouds are good bright calibration targets as they have a predictable reflectance. The amount
of cloud reflectance at the top of the atmosphere is reduced due to ozone absorption. DCC are at
the tropopause level and hence effects of water vapor and tropospheric aerosol absorption are
minimized. After correction of the ozone absorption DCC have almost a perfect white spectral
behaviour in VNIR bands and therefore the reflectance in one band can be extrapolated to an
other band to evaluate the accuracy of interband calibration coefficients. For PROBA-V , we will
implement a refinement of this method proposed by Lafrance et al. (2002) [LIT31] where TOA
reflectance over these clouds are modelled using radiative transfer calculations (RTC). This
makes it possible to take into account small spectral dependencies in the TOA cloud reflectance.
Figure 16: SCIAMACHY TOA reflectance spectra over DCC
The data used in Figure 16 has not been corrected for gaseous absorption, and has been prepared
at SCIAMACHY (courtesy of Dave Doelling).
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 77/139
4.5.2.2.1.3 Retrieval of the cloud optical thickness from the RED band
Because of the uncertainty in the absolute values of the TOA reflectance over DCC (see
variation in Figure 16) due to variations of cloud optical thickness cτ , DCC can’t be used for
absolute calibration . However due to the almost flat spectral signature DCC are adequate for
inter-band calibration , using a well-calibrated band as reference band to derive the cloud optical
thickness cτ from the image pixel itself by comparing image data with simulated data. The
cloud reflectance in the non-absorbing VNIR bands is mainly sensitive to the cloud optical
thickness (Figure 17), while in the absorbing SWIR bands it is more sensitive to the droplet
effective radius (Figure 16).
Figure 17: DCC TOA reflectance for in function of cloud optical depth.
Figure 18: DCC TOA reflectance in function of cloud effective radius.
Figure 17 and Figure 18 have been calculated for SZA=0°, SAA=40°, VZA=40° and VAA=30°.
In Figure 17 a cloud effective radius of 20 μm is assumed. For Figure 18 a cloud optical depth of
80 is used.
The method uses the TOA reflectance in the RED band corrected for ozone absorption to derive
the cloud optical thickness by inversion of a LUT of TOA reflectance in function of cloud optical
thickness assuming a fixed ice particle model.
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
10 60 110 160 210 260 310 360
Cloud Optical Depth
TO
A r
efl
ec
tan
ce BLUE
RED
NIR
SWIR
0
0.2
0.4
0.6
0.8
1
1.2
5 15 25 35 45 55 65
Reff
TO
A r
efl
ecta
nce
BLUE
RED
NIR
SWIR
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 78/139
4.5.2.2.1.4 Retrieval of the calibration coefficients for the BLUE and NIR bands
The cloud optical thickness estimated from the RED band is then used to model the TOA
reflectance in the other bands ( )(mod, BLUEelc
TOA , )(mod, NIRelc
TOA ) through a radiative transfer
LUT.
The change in calibration coefficients for the Blue and NIR band is then calculated as :
)(
)(mod,
,
BLUE
BLUE
A
AA
elc
TOA
measc
TOA
old
BLUE
new
BLUEBLUE
;
)(
)(mod,
,
NIR
NIR
A
AA
elc
TOA
measc
TOA
old
NIR
new
NIRNIR
4.5.2.2.2 Algorithm implementation
4.5.2.2.2.1 Generation of LUTs
The generation of LUTs of TOA reflectance above Deep Convective Clouds is performed with
LibRadtran RTC. LibRadtran has been successfully validated in several model intercomparison
campaigns and by direct comparison with observations. LibRadtran takes into account cloud
phase, micro-physical properties as well as a complete description of the background atmosphere
and surface. Both water and ice clouds models are included. The microphysical properties of
water clouds are converted to optical properties either according to the Hu and Stamnes (1993)
[LIT34] parameterization or by Mie calculations. For the optical properties of ice clouds
calculations from Baum et al. (2005) [LIT35] are used.
Table 17 gives the properties of the cloudy atmosphere and the surface that are used to create the
LUTs with LibRadtran. For the on-the-fly calculated LUT the minimal and maximal values of the
angles are determined from the image itself.
The four most common ice crystals shapes are: bullet rosettes, aggregates, hollow columns, and
plates. Each particle type is characterized by its single scattering albedo, the extinction cross-
section and the scattering phase function which determine the interaction of light with the cloud
particle. As a result, for a given cloud optical thickness and sun/viewing geometry the LUT will
therefore depend on the microphysical properties of the ice particles. The ice particle model from
Baum et al. (2005) [LIT35] is used to create the LUTs. The Baum model is also employed in the
MODIS operational ice cloud optical depth retrieval (Zhang et al.,2009 [LIT36] ) and by Sohn et
al. (2009) [LIT33] for calibration over DCC. The Baum model is based on the use of in-situ
observations of ice particle sizes and habits to compute optical properties for a realistic ensemble
of theoretical particles (Figure 19: Mixing scheme according to Baum et al (2005) [LIT35]).
As cloud reflectance for the VNIR bands is rather insensitive to the effective particle radius, the
effective radius is assumed fixed for the calibration. According to Sohn et al. 2009 [LIT33]
MODIS effective particle radii for DCCs from one month of data show a narrow distribution with
maximum frequency at 20 μm. LUT calculations will be performed for an effective radius of 20
μm.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 79/139
Figure 19: Mixing scheme according to Baum et al (2005)
As deep convective clouds are usually thicker than 10km, the cloud-top and cloudbase heights are
assumed to be 15 km and 1 km, respectively (Sohn et al., 2009) [LIT33].
To calculate the Rayleigh scattering the pressure profiles from the tropical profile are used.
Gaseous absorption is not taking into account in the LUT calculations as PROBA-V
measurements will be corrected for ozone absorption
The ocean reflectance is assumed lambertian with a chlorophyll content of 0.1 mg/m³.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 80/139
Parameter LUT
Atmospheric profile Tropical profile
Aerosol Model (from 0 to 2 km) Maritime profile
Stratospheric aerosol Background aerosols
Solar irradiance file Taken from 6SV
Aerosol Optical depth 0.08
Gaseous absorption No molecular absorption
Solar zenith angles From 10° to 40°
View zenith angles From 0° to 30° °
Relative azimuth angles From 0° to 180°
Surface albedo (ocean) BLUE: 0.02; RED: 0.002; NIR: 0
(lambertian)
RTF solver DISORT2
Cloud optical depth 20 values from 40 to 200
(logaritmically scaled)
Ice particle radius 20 μm
Ice cloud optical properties Baum detailed
Location ice cloud From 1 to 15 km
Water cloud layer NO
Location water cloud N/A
Water cloud optical properties N/A
Table 17: Properties of the cloudy atmosphere and surface for DCC LUTs
4.5.2.2.2.2 Processing steps
Conversion to TOA reflectance k
TOA
convert the raw DNs of the clouds calibration image to TOA radiance, following
Equation 4
convert TOA radiance to apparent TOA reflectance, following Equation 8
Correction for gaseous absorption
Correct for the gaseous transmittance of 3O (conversion of TOA to measc
TOA
, ) based on
predefined exponential variation with airmass and ozone content from ECMWF or
TOMS data or climatology. The LUTs of TOA reflectance are calculated without gaseous
absorption, therefore the PROBA-V measurements have to be corrected for gaseous
absorption. As for convective clouds observations water vapour and oxygen gases lie
below the cloud level, their absorption can be neglected. The ozone layer is however
located mainly above the clouds and therefore ozone absorption can’t be neglected.
Automatical selection of suitable clouds for deep convective clouds calibration (DCC)
Suitable deep convective clouds develop over subtropical warm oceans in inter-tropical
latitudes between 30°N and 30°S. In PROBA-V Calibration Plan [PVDOC-615 ] three
zones based on experience with SPOT-VGT and personal discussions with Dave
Doelling (NASA, expert in DCC) are selected for daily acquisition remove all pixels for
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 81/139
which solar zenith angles > 40° or viewing zenith angles > 30° to avoid possible
shadowing effects
use only pixels for which relative azimuth is between 30 and 150° to reject observations
corresponding to viewing geometries near the specular direction.
remove all pixels for which 8.0, cNIR
TOA (cloud threshold) to ensure that thick clouds are
selected
remove all pixels for which RMSE( 03.0)/)( ,, cNIR
pp
cNIR
TOA xNN to select only
homogenous clouds (the exact size expressed in number of pixels shall be evaluated
during commissioning phase based on real PROBA-V data over DCC)
remove central pixels if 8.0)(, notallxNN pp
cNIR
TOA to select only pixels with cloudy
backgrounds
Retrieval op Cloud optical thickness
For each pixel the cloud optical thickness(τcest
): the simulated )(mod, REDelc
TOA (in the
global LUT or the on-the-fly calculated LUT) are interpolated on the angles from each
pixel to obtain )(mod,
REDelc
TOAρ corresponding to the geometry of the observation. Then
),( ,
mod,
ic
elc
TOA RED and ),( 1,
mod,
ic
elc
TOA RED that surround )(, REDmeasc
TOA are found
and assuming a linear relationship the cloud optical thickness(τcest
) corresponding to
)(, REDmeasc
TOA is estimated.
Calculate BLUEA and NIRA
For each pixel find ),(mod, NIRBlueelc
TOA corresponding to τcest
: ),,(mod, est
c
elc
TOA NIRBlue
is derived from a linear interpolation between ),,(mod,
i
elc
TOA NIRBlue and
),,( 1
mod,
i
elc
TOA NIRBlue at the estimated cloud optical thickness(τcest
)
Calculate for each pixel the change in calibration coefficients for the BLUE and NIR
bands as :
)(
)(mod,
,
BLUE
BLUE
A
AA
elc
TOA
measc
TOA
old
BLUE
new
BLUEBLUE
;
)(
)(mod,
,
NIR
NIR
A
AA
elc
TOA
measc
TOA
old
NIR
new
NIRNIR
4.5.2.2.2.3 Required ancillary data
The only ‘external’ ancillary data needed in the processing are the total ozone amount to calculate
the ozone absorption and MODIS Terra Temperature images (or equivalent sensor if MODIS not
available) for proper cloud selection.
Other ‘internal’ ancillary data (assumed attached to the input image) : time, date, sun zenith
angles, solar zenith angles and relative azimuth angles.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 82/139
4.5.2.2.2.4 Summary of the processing steps
Figure 20: Flowchart DCC calibration
4.5.2.3 Error analysis
The error sources are :
1. Cloud effective radius: an error of 10 μm is considered in the error analysis
2. Microphysical properties model: Baum mixture of particles compared to pure crystal
shape clouds such as plates, hexagonal column, hexagonal crystals. Calibration is
performed with the Baum mixture of particles which may differ from the real
microphysical model The degree to which the calibration coefficient is dependant on the
model was estimated by carrying out the calibration process with a LUT calculated with
the Baum mixture on an artificial image generated using the plates model .
3. Cloud top height: the uncertainty in the cloud altitude leads to an uncertainty of the
Rayleigh scattering contribution above the clouds. In the error analysis a 3 km error in
cloud top altitude is considered.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 83/139
4. Cloud geometrical depth: an uncertainty of 4 km in the cloud geometrical depth is used
for the error analysis
5. Ozone: an error of 20 % in the ozone concentration is considered in the sensitivity
analysis
6. Atmospheric profile: uncertainties are assessed by using a mid-latitude summer profile
instead of the tropical profile
The error analysis is performed by using a LUT generated according to the specifications in Table
17. This LUT is then used to perform a DCC calibration on an artificial image generated
according to the same specification except for the parameter of interest (e.g. for case 1 cloud
effective radius: images were created with a cloud effective radius of 10 μm and 30 μm instead of
20 μm). In the error analysis we considered four different view zenith angles ( 0°, 10°, 20°, 30°),
six cloud optical thickness values (50, 60, 70, 80, 100, 120) and four different sun zenith angles
(0°, 10°, 20°, 30°).
Table 18 gives both the relative (interband) error budget and the absolute error for the DCC
method for both BLUE and NIR band. The table contains the average error, the error at 1-sigma
and at 2-sigma The error sources are assumed to be non-correlated and thus the total error is the
quadratic sum of all errors. For the absolute error the uncertainty in the RED band calibration is
also considered.
Error Sources
BLUE NIR
average
error
(%)
1-sigma
error
(%)
2-sigma
error
(%)
average
error
(%)
1-sigma
error
(%)
2-sigma
error
(%)
Cloud effective radius (10 μm) 0.300 0.531 0.761 0.886 1.124 1.363
Microphysical model 0.218 0.358 0.499 0.758 0.984 1.210
Cloud top height 0.095 0.133 0.172 0.011 0.018 0.024
Cloud geometrical depth 0.040 0.043 0.046 0.002 0.004 0.005
Ozone (20%) 0.359 0.379 0.399 0.397 0.429 0.460
Atmospheric profile 0.003 0.005 0.006 0.001 0.002 0.003
Red calibration error (3%) 2.912 2.929 2.945 2.704 2.882 3.061
Total interband 0.526 0.757 1.009 1.231 1.554 1.880
Total absolute 2.959 3.025 3.113 2.971 3.275 3.592
Table 18: Calibration error budget for DCC
The method inaccuracy is mostly due to uncertainty in the cloud effective radius, the
microphysical properties model and ozone content. The other uncertain parameters have only a
very small effect on the accuracy of the DCC calibration method. Interband calibration can be
performed over DCC with an accuracy of 1% (2-sigma) for the BLUE band and better than 2%
for the RED band.
In section 0 it is explained how the uncertainty calculation is used in the determining the final
error budget.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 84/139
4.5.3 Multi-temporal Calibration over stable deserts
The goal of multi-temporal calibration is to monitor the stability and variations of the sensors
responses over time. This results in information about the long-term behaviour of the sensor
which can be compared with the temporal variations in the coefficients determined with the
absolute calibration methods. Multi-temporal calibration can be achieved by regular monitoring
of targets that are stable over time. Due to the uncertainties related to the seasonal artifacts in the
deserts, at least almost one year of PROBA-V data are needed to get accurate multi-temporal
calibration results.
4.5.3.1 Introduction
The multi-temporal calibration of deserts doesn’t rely on extern BRDF data sources, but on the
PROBA-V desert data itselves. It assumes temporal stability of the surface, andlow variability of
overlying atmosphere. All new acquired data are compared to comparable data of the same desert
site in this reference database. Comparable means they have approximately corresponding
geometries. More specifically a PROBA-V desert archive dataset over a period of almost one
year. The angles defining the geometry are the sun zenith angle (SZA), the view zenith angle
(VZA) and the relative azimuth angle (RAA).
The comparison is performed on BOA reflectance, meaning that all data are atmospherically
corrected data with SMAC considering ozone, water vapour, pressure (from ECMWF) and AOT.
A large reference PROBA-V reference database is needed in order to find acquisitions fulfilling
this condition and to remove/correct for seasonal effects.
4.5.3.2 Algorithmic implementation
The approach is schematically given in Figure 21 and described in more detail in the next
paragraphs.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 85/139
Figure 21: Flowchart multi-temporal calibration over deserts
4.5.3.2.1 Generation of reference database
For a sufficiently large time period a database of PROBA-V BOA reflectance data over the
different deserts sites () needs to be compiled. For this, each Level1B image over a desert site has
to be processed in the following way:
a. Preprocessing of the L1B data to projected TOA reflectances:
Conversion to DNs to TOA radiance refmeas
TOAL , by applying current calibration
coefficients following Equation 4.
Conversion of TOA radiance refmeas
TOAL , to apparent TOA reflectance
refmeas
TOA
,
(Equation 8)
Projection of the the data
b. ROI extraction and remval of cloudy scenes according procedure described under 4.4.4.2.3
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 86/139
c. Averaging of refmeas
TOA
, over the scene, which gives refmeas
TOA
,
d. Atmospheric correction using SMAC, considering O3, H2O, Pressure, AOT (= 0.2) and the
desert aerosol model which results in the average bottom of atmosphere reflectance refmeas
BOA
,
e. Storage of refmeas
BOA
, together with the sun and viewing geometry for the observation
((SZAref, VZAref, RAAref) in reference database
4.5.3.2.2 Finding comparable acquisitions
First, for each new acquisition, the above described processing steps (a to d) have to be applied to
the data to retrieve the average bottom of atmosphere reflectance of the new acquisition newmeas
BOA
,
.
Next, the reference database is searched for comparable acquisitions over the desert site. Such a
pair consists of one PROBA-V and one PROBA-V reference observation of the same desert site
which are expected to show no differences in result induced by geometry.
Comparable means they have either approximately identical or approximately corresponding
geometries.
The angles defining the geometry are the sun zenith angle (SZA), the view zenith angle (VZA)
and the relative azimuth angle (RAA).
Identical geometry means that these 3 angles are the equal for both observations. Then the TOA
reflectance can be assumed the same:
TOA
REF(SZAREF, VZAREF, RAAREF) = TOANEW (SZA NEW, VZA NEW, RAANEW)
Corresponding angles are pairs of angles that yield the same reflectance due to the reciprocity
principle. This states that if SZA and VZA are exchanged, the same result is obtained:
TOA
REF(SZAREF, VZAREF, RAAREF) = TOANEW (VZANEW, SZANEW, RAANEW)
If we also assume that the reflectance behaves symmetrical with respect to the principal plane, we
also obtain corresponding angles for a change from + RAA tot -RAA:
In practice, pairs of angles will normally never match exactly. However, an exact match is also
not necessary: similar angles will still lead to comparable reflectances. Therefore we relax the
requirement to approximately comparable angles.
A angle pair is considered approximately matching if :
(SZAREF -SZANEW)2 + (VZAREF -VZANEW)
2 + (RAAREF -RAANEW)
2 /4 < tol
2
where a sensible value for the tolerance is: tol = 10
A angle pair is considered approximately matching if :
(SZAref –SZAnew)2 + (VZAref –VZAnew)
2 + (RAAref –RAAnew)
2 /4 < tol
2
where a sensible value for the tolerance is: tol = 10.
All reference acquisition fulfilling this conditions are averaged to retrieve the reference value ref
BOA .
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 87/139
4.5.3.2.3 Comparison of new acquisitions to its reference
Finally, an estimate of th the multi-temporal calibration coefficient is calculated as the ratio of newmeas
BOA
, to ref
BOA .
4.5.3.2.4 Outlier selection and daily averaging
A daily average is calculated by averaging the results obtained over the different sites after outlier
removal. A site is removed if it is detected as an outlier in at least one of the spectral bands. A
robust outlier selection procedure based on the median and standard deviation from the median is
used for this.
4.5.3.3 Error analysis
The main error sources are :
- Aerosol (model and AOT) variation between new acquisition and reference acquisition
- Stability of the site reflectance between new acquisition and reference acquisition due to
stability of the site itself or due to small difference in sun and view geometry
- Small errors may arise due to uncertainties in the meteo data used to perform the
atmospheric correction.
A quantification of the uncertainty has been performed by Hagolle and Cabot, 2002 [LIT51] and
Hagolle and Cabot, 2003 [LIT52]; The uncertainty (expressed as RMSE) ranged from 3.4 % for
blue bands to 1.8 % for NIR bands.These errors are however reduced by averaging over sites and
time.
Prototype activities an SPOT-VGT (see PVDOC-647) have indicated that by averaging over sites
and time long term multi-temporal calibration uncertainties less than 1% can be achieved for
RED, NIR and SWIR bands.
4.5.3.3.1 Interpretation of long term time series
The coefficients obtained per day are in turn used in the statistical trending analysis (see section
4.6.2.3). This analysis uses statistical procedures to derive the best estimate for the current
coefficients, based on present and previous data. It includes a short term linear model of the
evolution of the coefficients, which is used as an estimation tool.
The long term evolution of the coefficients however merits a separate investigation to aid our
understanding of the process and the underlying physics. We plot the results over a sufficiently
long time period (> 1 year). An example for SPOT-VGT using monthly averages is shown in
Figure 22. Seasonal variations are clear from the plot, but also the longer term downward trend is
obvious.
That trend, as expected, appears as an exponential decay. It is possible to describe this by fitting a
logarithmic model to the data using classical least squares minimization:
A (t) = a.ln(t) + b.t + c
The obtained curve, (also shown for the example in Figure 22), represents the overall sensitivity
decay, with short term variations and seasonal effects filtered out. The decay curve can be
computed regularly and so the evolution of the model can also be monitored. The obtained trend
will correspond to the results of absolute calibration curve. It also permits to calculate indirectly
a calibration coefficient, or even predict it for coming months.
As the decay curve of PROBA-V is expected to be different from these of SPOT-VGT, decay
curves derived on the basis of SPOT-VGT can’t be used for PROBA-V. Furhermore it can be
noticed that the slope of the decay curve of SPOT-VGT2 is much steeper in the months after
launch that several monhts later. All this implicates that a long time series of PROBA-V data (>
1 year) is needed before sufficiently accurate decay curves on the basis of deserts can be derived.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 88/139
Figure 22: example of multi-temporal variation (taken from SPOT-VGT2)
4.5.4 Multi-temporal Antarctica
4.5.4.1 Introduction
Multi-temporal calibration over Antarctica has the same goal than that over desert sites, to
monitor the stability of the sensors response over time.
4.5.4.1.1 Advantages
The use of test site on Antarctica has several advantages compared to desert test site, we list them
here:
- very uniform targets
- very stable targets (also very bright in VNIR)
- high accessibility: there are 6 or 7 overpasses, so several images can be acquired in 1 day.
- high altitude sites (no aerosol)
4.5.4.1.2 Disadvantages
However, it also as some disadvantages:
- only possible during summer months at Antarctica (December/January)
- special acquisition required
- necessary to change integration time change to prevent saturation
- BRDF effects for snow and ice are substantial, proper BRDF model is needed if observations
are made under a different sun-viewing geometry
Because of all the previous, the use of a test site on Antarctica is a valuable addition to the
calibration plan, especially to validate the results of the calibration over deserts (multi-temporal
and cross sensor calibration)
4.5.4.1.3 Calibration area
We will use the well known “Dome C” area, its location on Antarctica is shown in Figure 23. It is
about 716 x 716 km² large. The area is divided into 36 grid boxes of 120 x 120 km². Dome C
station is located in grid box 15. The calibration is done on 4 of these grid boxes where spatial
studies have indicated that the area within grid box is sufficient homogenous.
VGT2/B2 calibration model fit in June 2006
0,82
0,84
0,86
0,88
0,90
0,92
0,94
0,96
0,98
mai-02 nov.-02 mai-03 nov.-03 mai-04 nov.-04 mai-05 nov.-05 mai-06
month average cal. june06
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 89/139
Figure 23: Location of Dome C calibration site (75°s, 123°E)
4.5.4.2 Algorithm
4.5.4.2.1 Data collection
Antarctica is not part of the normal operational imaging of PROBA-V. This means that no daily
routine acquisitions are available. Calibration can only be achieved from a dedicated acquisitions.
Dedicated acquisitions are needed also for another reason: the targets observed are so bright that
they would saturate under normal operational settings. Therefore special settings are necessary:
Images should be taken with low integration time settings.
4.5.4.2.2 Data processing
Data processing can be performed in the same way as for the multi-temporal calibration over
desert sites ( using BRDF and aerosol conditions relevant for Antarctica).
4.5.4.3 Error analysis
Earlier studies by CNES reported accuracy in the range of 2 to 3%.
4.5.5 Multi-temporal Moon
4.5.5.1 Introduction
The surface of the moon is stable over thousands of years. Although brightness of the moon
varies with illumination and viewing geometry, the photometric stability of the lunar surface
enables to characterise these cyclic variations with high precision. For these reasons the CEOS
WGCV has endorsed the Moon as a reference standard for calibration stability (Stone, 2008
[LIT40]).
Through a pitch manoeuvre lunar calibration images will be acquired once per month at a phase
angle near 7 degrees. This phase angle is preferred as it maximizes the illuminated surface of the
moon while avoiding increased uncertainty by minimizing the opposition effect. For each lunar
calibration the radiances observed by PROBA-V are integrated over the lunar image images.
Although the surface of the moon remains unchanged over the lifetime of PROBA-V, the
observed radiance varies due to :
Variations in Sun-Moon distance
Variations in PROBA-V-Moon distance
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 90/139
The phase angle differences
The libration angle differences
Variation in oversampling of the lunar images in the along track direction
The complexity of these dependencies mandates the use of a lunar radiometric model. The
USGS lunar calibration program has developed such an operational lunar model based on more
than 85000 images acquired by the ground-based Robotic Lunar Observatory (ROLO). An
empirical derived analytical model has been fitted through all these ROLO images (which are
corrected to TOA radiance, spatially integrated to irradiance, converted to standard distances
and finaly converted to effective disk reflectance). The atmospheric extinction correction is
performed by dedicated stellar observations. The model uses the star Vega and published stellar
flux data for absolute scale.
The ROLO lunar disk reflectance model is given as
where Ak the disk-equivalent reflectance, g the absolute phase angle, θ and ϕ the selenographic
latitude and longitude of the observer and Φ the selenographic longitude of the Sun. The model
coefficients are given in Kieffer and Stone (2005).
4.5.5.2 Procedure
Satellite image processing
The raw lunar images (in DN) are first calibrated to radiometric quantities. Then the integrated
irradiance attributed to the Moon is determined by summing all pixels on the lunar disk including
the unlit portion. A critical aspect here is the determination of the lunar disk, which might be a
bit subjective. Different techniques are published in literature. In Stone et al. (2013) [LIT58] a
threshold technique, combined with a proximity test, is used to select the on-Moon pixels.
Another techniques exists in first finding the lunar limb and then fitting an ellipse through the
points on this limb ([LIT57]). The points of the lunar limb can be determined as the inflection
points (where second derivative is zero) in the rise of brightness onto the moon at the lit edge.
Once all pixels on the moon are determined their radiances are summed and multiplied by the
solid angle ( subtended by one pixel :
∑
The solid angle ( subtended by one pixel is computed as :
with and the nominal angular size of a pixel in respectively X and Y direction.
Geometric processing
Next the oversampling factors need to be determined. The oversampling factor ( can be
determined either from spacecraft attitude data or from the apparent size of the moon in the lunar
image. Finally a correction ( for the actual Sun-Moon distance and the instrument-Moon
distance has to be performed with calculated as :
(
)
(
)
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 91/139
with MLD the mean Earth-Moon distance (384400 km).
The instrument-lunar irradiance at standard distance is :
These distances can be calculated based on the selenographic coordinates of the Moon, Sun and
spacecraft, calculated on the basis of high precision ephemeris.
Comparison with ROLO Model
As the ROLO model works with reflectances instead of irradiances a conversion of the
instrument-lunar irradiance to effective disk reflectance is needed :
With the solid angle of the Moon (=6.4177 x 10-5
sr) and the solar irradiance.
Inputs to the ROLO model are : the absolute phase angle, the selenographic latitude and longitude
of the observer and the selenographic longitude of the Sun. These can be calculated on the basis
of the ephemeris time of the Moon observation and the instrument location at this time.
On the basis of these parameters the ROLO modeled disk-equivalent reflectance
can be calculated (and spectral interpolating, resampling with instrument specific response
curves) and compared to the measured reflectance. The relative difference is given as :
The monthly measurements/model comparisons are then used in a time series analysis in order to
track the radiometric stability.
4.5.5.3 Performance evaluation
The achievable accuracy of the lunar calibration depends on the lunar irradiance model
predictive precision , underpinned by the inherent stability of the Moon and the random errors
associated with calculation of the irradiance values from the PROBA-V lunar observations
(Stone, 2008 [LIT40]).
The USGS lunar irradiance model fits the ROLO observations with an average residual over all
bands just under 1%. This value effectively represents the ‘relative’ precision of the model’s
predictive capability (Stone, 2008 [LIT40]). Stability monitoring can be achieved with sub-
percent per year precision.
The main advantage of lunar multi-temporal calibration over desert sites is that calibration results
can obtained immediately while for the desert calibration data over at least one year are needed to
correct for seasonal variations.
Due to uncertainties in the ‘absolute’ lunar irradiances, the moon can not be used as an absolute
radiometric standard.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 92/139
4.6 Statistical analysis on calibration coefficients
4.6.1 Introduction
Let us assume that we have (different operational methods to estimate the absolute calibration
coefficient ( kA ) for the different spectral bands (k). The operational methods in use are currently:
Rayleigh: estimate kA for BLUE, RED, needs NIRA as input
Deserts: estimate kA for all bands: BLUE, RED, NIR and SWIR
Sun Glint: estimate kA for BLUE, NIR and SWIR, needs REDA as input
Clouds estimate kA for BLUE and NIR, needs REDA as input
For each method PROBA-V images are acquired over a number of sites during several
acquisition days. For all the suitable pixels in the image an absolute calibration coefficient is
calculated. The topic we address here is how to calculate the best estimate of kA at the current
day ct based on all these different estimates of kA ?
The approach to calculate the best estimate of kA is graphically illustrated in Figure 24 and
explained in detail in the following sections. For statistical analysis we follow roughly the
approach of Govaerts et al., 2004 [LIT41] . The trend analysis and the final reconciliation of the
different methods are similar to the MISR approach (Bruegge et al., 1999 [LIT42] ). In the next
paragraph we will omit the index k for the spectral band.
Figure 24: Calculation of best estimate of kA
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 93/139
The following notations are used in this chapter (Table 19):
t Time index expressed in days
m Method index
g Site index
p Pixel index
P Number of pixels
M Number of methods
G Number of sites
T Number of acquisition days
Table 19 : Notations
4.6.2 Approach
4.6.2.1 An individual image (1 method, 1 site, 1 acquisition day)
First, we consider only one site of one method at one acquisition day. Following the calibration
method specific pre-processing steps P ‘suitable’ calibration pixels are selected in the image. For
all these suitable pixels an absolute calibration parameter is calculated:
)(,, pA gtm with p from 1 to P
4.6.2.1.1 Outlier detection and handling
A statistical analysis is performed on the pN absolute calibration coefficients to remove outlier
values. This outlier selection method uses the median (MED) and the Median Absolute Deviation
(MAD) instead of the mean and standard deviation. While the mean and standard deviation can
be affected by a few extreme values or by even a single extreme value, the median and MAD are
minimally effected by the outliers. gtmMED ,, and tmgMAD ,, are calculated as
Ppgtmgtm pAmedianMED
...1,,,, )(
and
Ppgtmgtmgtm MEDpAmedianMAD
...1,,,,,, |)(|
When the MAD value is scaled by a factor of 1.483, it is similar to the standard deviation in a
normal distribution:
gtmMADgtmMADS,,483.1,,
Pixels with an absolute calibration value outside the 3-sigma confidence interval are removed.
The selected pixels have therefore to obey the following condition:
gtmMADgtmtmggtmMADgtm SMEDpASMED,,,,,,,,,, 56.2)(56.2
After removing the outliers the site-averaged calibration parameter is calculated for the
remaining Pc values as
cP
p
gtm
pc
gtm pAN
A1
,,,,
1
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 94/139
4.6.2.1.2 Error assessment
This associated relative error )( ,, tmgA is calculated as the square root of the quadratic sum of
all error contributions which can be assumed uncorrelated. These error contributions are :
1. Errors due to uncertainties in atmospheric parameters (atm ) (e.g. the relative uncertainty
introduced in the absolute calibration coefficient due to for instance the uncertainty in
meteo-data)
2. Errors due to uncertainties in surface characterization ( surf ) (e.g. uncertainty in
reference BRDF)
3. Errors due to ‘calibration errors’ in reference bands (cal )
4. Errors due to intrinsic uncertainties in RTF model (6SV, Libradtran,modtran) ( rtf )
5. Errors due to noise (due to both nstrument noise and target non-uniformity) ( noise )
Therefore the total relative error for tmgA ,, can be written as :
2
,,
2222
,, )(gtmnoisemrtfmcalmsurfmatmgtmA
The first three errors ( atmε , surfε , calε ) are included in the error analysis performed for the
different calibration methods. A distinction can be made between random and systematic errors.
Random errors ( atmε , surfε ) can be reduced by spatially averaging over several sites (see next
section).
gtmnoise ,, is estimated from the image at a confidence interval of 95 % and assuming a normal
distribution as :
cgtm
gtmA
gtmnoisePA
S
,,
,,
,,
96.1
with
2
1
,,,,,,))(
1
1
cP
p
gtmgtm
cgtmA
ApAP
S
This analysis results in an average absolute calibration parameter gtmA ,, and associated relative
error )( ,, gtmA calculated based one site, from one method and one acquisition day.
4.6.2.2 Spatial averaging (all sites for one method, one acquisition day)
Calibration coefficients derived over all calibration sites for one method and one moment in time
are spatially averaged to reduce random errors . These random errors can be reduced by the
square root of the number of sites assuming that these errors are not correlated in space.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 95/139
Figure 25: Spatial averaging
A robust outlier detection test using median and MADS estimates is performed similar to the one
in section 4.6.2.1.1. Sites with calibration coefficients that fall outside the confidence interval are
assumed outliers and are disregarded.
When a site is often indicated as an outlier, it should be investigated if this systematic bias is due
to for instance an erroneous surface characterization (eg. marine reflectance not valid for that
site).
Once the outliers are removed the weighted mean tmwA,
over the Gc remaining sites is
calculated:
Gc
g
gtmgtmtmw AwA1
,,,,,
The weights for each site calculated using normalized weights inversely proportional to the site-
average calibration error )( ,, gtmA as :
Gc
g gtm
gtm
gtm
A
Aw
1
2
,,
2
,,
,,
)(
1
)(
1
The associated relative error )( ,tmA at 2-sigma is calculated as :
222
,
22
,GcGc
)(mrtfmcaltmnoise
msurfmatm
tmA
with
tmw
tmwAGc
tmnoiseAGc
St
,
,1,2
,1
.
and tmwA
S,
the standard deviation of the weighted mean .
4.6.2.3 Trending analysis (all sites for one method, different acquisition days)
The greatest change in sensor response is expected immediately following launch. However after
a few months in orbit the responsivity of each camera is expected to change more slowly in time.
The characterisation of this responsivity change require a large number of observations over time
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 96/139
due to the uncertainty with which a single observation is made. A trending analysis is performed
on the time series of coefficients obtained per method. Each newly generated coefficienttmwA,
will be combined with previous generated coefficients and the best trend line is calculated. This
procedure is explained in this paragraph. The important things to be determined in the trending
analysis are: (1) the trending model, (2) time scale, (3) the weighting of the different
observations.
4.6.2.3.1 Linear Model
We choose a linear model. Then we can use linear regression, which is performed using the
classical least squares optimisation. For every acquisition day t we have calculated absolute
calibration coefficient tmwA,
. The set of T (t, tmwA,
) pairs for the time interval [t+1-T,t] is used to
fit a linear model
tbaAModelm,twtTt 0...1 )(
The offset a0 and the slope b of the model are computed from the data so that the sum of the
squared difference between t,mwA and their model values )(...1 m,twtTt AModel
is minimized
or using the notation:
t
Tt
XT
X1
1
We can express the solution of the linear regression as:
22
.
tt
tAtAb
m,twm,tw
The result of the linear regression is shown as the middle red line in Figure 26.
Figure 26: Result of linear regression, with 1 and 2 sigma confidence intervals
4.6.2.3.2 Linear Weighted Model
The previous model assumed equal importance of all data points. However we want to assign
weights to individual data points. This has two reasons:
every data point was determined with a different accuracy )( ,tmA ; for simplicity
denoted as εt. These are accounted for by weighting with a factor : 1/εt2
0,78
0,8
0,82
0,84
0,86
0,88
0,9
0,92
0,94
0 2 4 6 8 10
data points
model
lower
upper
lower2
upper2
tbtAam,tw0
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 97/139
the data points closer to the current time are assumed to be more important than the
earlier points. We propose a time window with weights that linearly increase from 0 at
the starting time tstart to 1 at the end time tstop. :
startstop
start
tt
tw
''
Figure 27: time window with linearly increasing weights
For the time window of size T, with tstop = t and tstart = t+1-T, we obtain:
1
)1(''
T
Ttw
Both weights can be combined into a single weighting factor:
2)1(
)1('
T
Ttw
If we divide by the average weight <w’>, we obtain weights which sum equal to T:
'w
τ'wτw
The weights are taken into account in the linear regression model by using weighted averages:
t
Tt
XwT
wX1
1
Using these we derive:
2wt2wt
wt.m,twAwt
m,twAw
wb
22
2
0
..
wtwt
tAwwtwtAwa m,tm,t ww
w
The results of weighted linear regression are shown as the middle red lines in Figure 28 and
Figure 29. In Figure 28, only temporal weights are used. Compared to the previous result, this
follows more closely the last (more recent) points in time.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 98/139
Figure 28: Result of linear regression, including temporal weighting.
Figure 29: Result of linear regression, including temporal weighting and individual
accuracies
4.6.2.3.3 Accuracy of the unweighted model
The standard deviation about the regression ( rS ) is given by
)2(
2
T
SbSS TTAA
r
with
22
m,twm,twAA AATS , 22 ttTSTT
The standard deviation for the predicted or modelled )(...1 m,twtTt AModel value at time τ is
given by
0,78
0,8
0,82
0,84
0,86
0,88
0,9
0,92
0,94
0 2 4 6 8 10
data points
model
lower
upper
lower2
upper2
0,78
0,8
0,82
0,84
0,86
0,88
0,9
0,92
0,94
0 2 4 6 8 10
data points
model
lower
upper
lower2
upper2
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 99/139
TT
rwAS
t
TSS
21
1)(
The 95 % confidence interval on the modelled value )(...1 m,twtTt AModel is then given by +/-
1.96 )(tS A
Logically, the model will yield the smallest errors for values of τ close to the average value <t>
used to generate the model. Further away, the error will be larger. The confidence intervals are
also plotted for the example in Figure 26. The intervals are small in the middle of the
observations, and larger near the sides.
4.6.2.3.4 Accuracy of the weighted model
For the weighted regression the formulae become:
)2(
2
T
SbSS TTwwAAw
rw
with
22
m,twm,twAAw AwAwTS , 22 wtwtTTTwS
The standard deviation for the predicted or modelled )(...1 m,twtTt AModel value for the
weighted model at time τ is given by
TTw
rwAwS
wt
TSS
21
1)(
The 95 % confidence interval on the modelled value )(...1 m,twtTt AModel is then given by +/-
1.96 )(AwS
The last equation was use to determine the confidence intervals in Figure 28 and Figure 29. The
intervals are smaller near the end. In Figure 29, individual accuracies are given: some points get
a larger error and are used less in the regression. As a result; the confidence intervals become
narrower.
4.6.2.3.5 Result of the trending analysis
Every day t, a new coefficient )(...1 m,twtTt AModel is obtained. The model from the previous
day gives the best estimate for the coefficient of the present day, without using actual value of
that day t,mwA :
tbaAModel tTttTtm,twtTt 1...1...,01... )(
We can also determine a (99%) confidence interval around the estimated value by adding
)()56,2( tS A
We mark )(...1 m,twtTt AModel as a potential outlier if it falls outside the confidence interval. It
is not used in further models, unless it is unmarked at a later moment in time.
Every day, the new estimate )(...1 m,twtTt AModel is computed from the model by adjusting the
time window. The new estimates together constitute the time evolution. Values marked as
potentials outliers are not used in the model. After the model is set, all potential outliers are
rechecked against the criterion. If they now fall into the interval they are unmarked and then no
longer treated as outliers.
The uncertainty in )m,twA(t...T1tτModel (at 2-sigma) is calculated as follows :
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 100/139
2
,
2
,
2
,
...1 )( fitmsystematicm
randomm
m,twtTtT
AModel
with )(
)(96.1
...1
,
m,twtTt
Afitm
AModel
tS
Note that the random and systematic errors are treated separately, as explained in the previous
section. It is important that a critical assessment is made to estimate systematic errors (bias) for
every method. E.g. if a method (e.g. desert calibration) shows substantially different values than
other methods in other calibration systems, it is likely that this method contains some unknown
source of systematic error.
4.6.2.4 Combining different methods
The previous analyses are performed for each calibration method separately. Now the calibration
coefficient from the trending analysis of the M different methods will be combined in a single
calibration parameter. Not all methods yield coefficients for all different bands.
4.6.2.5 Combined coefficients: combining function Fcomb
The )(...1 m,twtTt AModel coefficients from the M different methods are combined in a single
twA for day t. The is done by computing the weighted average as :
M
mm,twtTttmtw AModelwA
1
...1, )(
with
M
m wtTt
wtTt
tm
m,t
m,t
AModel
AModelw
1
2
...1
2
...1
,
)(
1
)(
1
This combining function will be denoted in the next section as Fcomb.The use of a weighted
average ensures that the more accurate results contribute more to the final result.
4.6.2.5.1 Combined accuracy
The relative uncertainty of the combined coefficienttwA should reflect the individual
uncertainties )(...1 m,twtTt AModel , which include systematic errors. However, even if
systematic errors for individual methods have been assessed, this does not guarantee that all bias
between methods has been covered.
From the collection of individual coefficients tmwA,, we can make an estimate of the in-between
method bias. Since only a few coefficients are available, estimating the bias in the classical
statistical way does not lead to useful results. Instead we use the method as proposed in Levenson
et al 2000 [LIT43]. This assumes a reasonable value for the bias, and describes the bias with a
rectangular distribution with extreme values ± b, which has a standard uncertainty of = bm/√3.
A sensible assumption is that the in-between method bias equals the largest difference between
the values for individual methods. We then obtain the following for the uncertainty caused by the
in between method bias:
)()(3
1)( ...1:...1:
2
)...1:( m,twMmm,twMmm,twMminbetween AMINAMAXA
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 101/139
The combined accuracy can be written as the weighted average of the individual uncertainties
plus the in between method bias:
M
mm,twMminbetweenm,twtTttmt AAModelw
1
2
)...1:(...1, )(²)(
Taking into account the definition of wm,t, we can rewrite this to:
)(
)(
1
2
)...1:(
1
2
...1
m,t
m,t
wMminbetweenM
m wtTt
t A
AModel
M
4.6.2.5.2 Interband dependencies
An additional difficulty arises: some methods (Rayleigh, Sun glint) need the value of the
coefficient of one band to estimate the other bands. This causes following dependencies:
input NIRA : Rayleigh method yields estimates: )( NIRBLUE
Rayl AA , )( NIRRED
Rayl AA
no input: Deserts method yields estimates: BLUE
DesA , RED
DesA , NIR
DesA , SWIR
DesA
input REDA : Sunglint method yields estimates: )( REDBLUE
Gl AA , )( REDNIR
Gl AA , )( REDSWIR
Gl AA
input REDA : Cloud method yields estimates: )( REDBLUE
Cl AA , )( REDNIR
Cl AA
Only the desert method works without input from other estimates. Since simultaneous estimation
using different methods is not practically feasible, we propose following flow:
Assume that coefficients have been reliably estimated at time t-1, by the combining function Fcomb that uses the estimates obtained from the different methods (omitting the notation
t...T1tτModel ) :
)1(),1(),1(),1()1( tAtAtAtAFtA BLUE
Cl
BLUE
Gl
BLUE
Des
BLUE
Raylcomb
BLUE
)1(),1()1( tAtAFtA RED
Des
RED
Raylcomb
RED
)1(),1(),1()1( tAtAtAFtA NIR
Cl
NIR
Gl
NIR
Descomb
NIR
)1(),1()1( tAtAFtA SWIR
Gl
SWIR
Descomb
SWIR
Due to the dependencies, the estimates at time t can be obtained only if we use initial estimates
)(, tA initRED , )(, tA initNIR . Several option can be considered:
1. The simplest solution is to take:
)1()(, tAtA REDinitRED
)1()(, tAtA NIRinitNIR
This ignores the fact that the desert method already yields results at time t.
2. We can incorporate these results by writing:
)(),1()(, tAtAFtA RED
Des
RED
Raylcomb
initRED
)(),1(),1()(, tAtAtAFtA NIR
Des
NIR
Gl
NIR
Clcomb
initNIR
3. A more advanced approach would be to predict the change in the other methods, e.g.:
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 102/139
)1()()( tAtAtA RED
Rayl
RED
Rayl
RED
Rayl
by assuming that it is proportional to the change in the desert method
)1()()( tAtAtA RED
Des
RED
Des
RED
Des
Then we predict:
)1(
)1()()1()()1()(,
tA
tAtAtAtAtAtA
RED
Des
RED
RaylRED
Des
RED
Rayl
RED
Rayl
RED
Rayl
initRED
Rayl
Which then leads to:
)(),()( ,, tAtAFtA RED
Des
initRED
Raylcomb
initRED
The other coefficients can be handled in the same way. The initial estimates can be used as input into the applicable methods in order to obtain first
iteration estimates ( )(1, tARED ) for time t. In general the estimates are not exactly equal to the
initial estimates: )()( ,1, tAtA initREDRED , …
Therefore we can iterate the calculations:
)(),()( ,, tAtAFtA RED
Des
initRED
Raylcomb
initRED
This iteration can be repeated until the values converge. In practice, this should happen after only
a few iterations. If this is not the case, it points to a severe problem. Then, the coefficient of the
previous day is used, while the cause of the problem should be investigated by an expert.
4.6.3 Updating the operational coefficient
The true calibration coefficients are expected to vary only slowly over time. A new estimate of
them is obtained daily. The estimates show small variations from day to day. Only the variations
caused by true physical changes should result in updates of the coefficients. Variation caused by
the limited accuracy of the estimates should be ignored. This raises the question when operational
coefficients need to be updated.
4.6.3.1.1 Update Strategy
We propose to use an update strategy scheme in which an update is possible daily, but is only
performed if a significance criterion is satisfied (see below) Such a scheme is is preferably applied
over fixed update schemes because:
o compared to a daily update scheme, many unnecessary updates and possible spurious
fluctuations are thus omitted.
o compared to update schemes with longer intervals the reaction time to real changes is shorter.
This also avoids that the choice of the intervals influences the results (a change occurring in
the middle of a month should be treated the same as the same change occurring at the end of a
month).
4.6.3.1.2 Significance Criterion
It is best practice to update the coefficients only when it makes sense. This notion is translated
into a significance criterion. A(t), the new best estimate of the coefficient at time t, is compared to
the coefficient at the previous update A(tupdate). The difference between the two is regarded as
significant if we are reasonably confident that the update reflects a real change and is not a
variation due to the estimation.
This is evaluated using a statistical test. Assuming large sample size, we can use the classical z-
test for two independent samples. Using the relative errors (t) and (tupdate) as accuracies, we
obtain:
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 103/139
)().()().(.96,1
)()(
2222
updateupdate
update
tAttAt
tAtAz
The z-score must be evaluated by comparing it against its critical value at the suitable confidence
level.. For high confidence levels (e.g. 95% or 99%), it answers the question “are we almost
certain that the values are different”.
In our case, if it is unclear whether the values are different or not, we do not want to assume that
they are the same, and therefore we still want to update. Thus in fact we have to evaluate the
opposite: “are we confident that the values are the same”. This is obtained by using a z-score
critical value for a lower confidence level (e.g. at 5 %, 20 % or even 50 %).
Figure 30: Values for which the difference is just significant at the given confidence level
This concept is illustrated in Figure 30: For a previous coefficient of 0,9, various new values are
plotted that correspond to the critical value at various confidence levels. For a chosen confidence
level, the value at the plot will be the smallest difference which will be seen as significant.
The choice of the confidence level is critical in this process. Its optimal value will be determined
based on calibration experiments, and will be monitored and optimized during commissioning.
4.6.3.1.3 Post monitoring of operational coefficients
The coefficients at time t are also estimated at time t. However when more time has passed, a more
accurate estimation can be performed. This is clear from the linear regression graphs in Section
4.6.2.3. Logically,the uncertainty about the linear model is smallest near the middle of the points
used. Some cases where future data clearly helps the estimation are:
if calibration is (partially) missing at time t, but available at time t - 1 and t + 1,
interpolating between them is more accurate than using only the results at time t-1.
If unusual variation is observed at time t, the processing will detect this as an outlier and
ignore it. Future data will confirm the data as an outlier or as a new trend, in the latter case
including the result is more accurate.
To test if the temporal model is performing well, it is good practice to recalculate the coefficients
for time t at a later time so that t lies in the middle of the temporal interval. Both estimations can be
plotted together for an additional monitoring of the time evolution estimates.
0,89
0,895
0,9
0,905
0,91
0,915
0,92
0,925
0,93
0,935
0,94
0 20 40 60 80 100
Co
eff
icie
nt
valu
es
Confidence Level
new coefficients
reference coefficient
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 104/139
4.6.4 Dealing with season-dependent errors
Both CNES and Govaerts and Clerici 2004 [LIT1] reported a seasonal trend in desert calibration
data. If these seasonal trends are not corrected for it can give erroneous results. These seasonal
trends will become only visible when we have more than one year of PROBA-V data. Seasonal
trends can be analysed with autocorrelation correlograms.
These seasonal effects will be analyzed with the SPOT VGT data over deserts from different
years. If these seasonal trends are consistent over the years, the seasonal trend curve will be used
to correct the desert calibration coefficients during the first year in orbit in order to minimize
sensor-independent seasonal variation. After one year in orbit these seasonal trend curves can be
replaced by PROBA-V derived curves.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 105/139
4.7 Camera to camera calibration
4.7.1 Introduction
In previous chapters the three different cameras are treated separately. This may result in biases
between cameras. To minimize inter-camera deviations a camera to camera calibration method is
used based on the overlap area between two adjacent cameras. This method allows also to
transfer calibration coefficients from the centrr camera to the outer cameras. This may be needed
to calibrate the SWIR band of the western looking camera for which the sun glint spot can’t be
observed.
PROBA-V has 3 cameras in order to fit the swath requirement. The field of view (FOV) of every
camera is 34.6 degrees. Since the outer cameras are pointed off-nadir for about 34.0 degrees,
there is a zone of overlap (Figure 31). The camera to camera radiometric calibration method
exploits the information delivered in the camera overlap zone. Since pixels in this zone are seen
by different cameras but with the same solar azimuth, solar zenith, view azimuth, view zenith and
approximately at the same recording time, BRDF effects (atmospheric BRDF, topographic BRDF
and target BRDF) have no influence in this calibration.
The camera to camera radiometric calibration methods will be used for:
(a) Continuous checks in the IQC using the "overlap-intermediate-products" delivered by the PF
to the IQC. Here, we continuously perform a regression between the TOA radiances of one
camera against the TOA radiances of the adjacent camera and perform (1) hypotheses tests to
detect if one or another camera starts to drift with respect to radiometry and (2) produce time-
dependent plots of the bias. This potentially enables the detection of time dependent effects due
to the orbital motion and/or seasonal effects. As such, these checks can deliver a trigger to the
IQC to actualize the radiometric calibration parameters supported by the other radiometric
calibration results.
(b) Discrete checks in the IQC: if one or another radiometric calibration method delivers a good
calibration result for a certain spectral band for a certain camera, this result can be transposed to
the other cameras using the information delivered by the camera-to-camera tie-points.
Camera to camera calibration uses the measurements of a limited part of the camera (the overlap
pixels) to draw conclusions about the calibration coefficients for the whole camera. Therefore it
critically depends on the validity of the equalisation coefficients of the individual pixels.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 106/139
Figure 31: Overview of the viewing geometry for one scan line.
In Figure 31, the settings are: platform position X, Y, Z, roll, pitch, yaw: 0.0°, 35.0°, 820km, 0°,
0°, -8.69°. FOV of one camera equals 34.6°. Outer cameras are pointed off-nadir for 34.0°.
4.7.2 Algorithm
4.7.2.1 Continuous checks
The DNs in overlap-intermediate-products (raw geometry) are first converted to TOA radiances
following Equation 4.
Next, the data are projected (see Figure 32) using a distance weighted interpolation technique
because this technique produces the least errors (Galbraith et al [LIT44]).
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 107/139
Figure 32 Camera overlap zone
An ASCII file containing TOA radiances for the overlapping pixels is produced using a mosaic-
generation application and for each band separately a linear regression between the TOA
radiances of one camera against the TOA radiances of the adjacent camera is then performed
(Figure 33).
If measurements are comparable all points will be scattered about the 1:1 line. The slope of the
regression is close to 1.0 and an intercept close to 0.0.
An alternative way of presenting the results is plotting the differences in TOA radiances between
pairs of overlapping points against their mean (Figure 34). This is called the Bland-Altman plot.
If the measurements are comparable, the difference should be small centred around 0 without
systematic variation with the mean value of pair of pixels.
2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.548.85
48.9
48.95
49
49.05
49.1
49.15
49.2
Center camera
Left camera
Camera overlap area
longitude [WGS84, degrees]
lati
tude [
WG
S84
, deg
rees
]
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 108/139
Figure 33: Regression of TOA radiances camera 1 versus camera 2
Figure 34: Regression of the differences in TOA radiances against their mean
These graphs are supplemented with formal statistical analysis (t-test for paired samples). Under
the circumstances that there is no obvious relation between the difference and the mean,
agreement between the TOA radiance values of both cameras can be tested with the following
statistics.
First, the mean relative difference d and the standard deviation S is calculated as
N
LLL
d
kcamTOA
N
kcamTOA
k
camTOA)/(
1,..1
1,2,
and
2
..11,1,2,
))/(1
1
N
kcamTOA
kcamTOA
k
camTOAdLLL
NS
with N the number of overlapping pair of pixels.
The 95 % confidence interval of agreement for individual differences between two cameras is
given by :
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 109/139
Sd 96.1
Finally, the 95% confidence interval for the bias is calculated as :
N
Sd
2
96.1
If this confidence interval doesn’t contain the value ‘0.0’, there consists a significant bias
between the two cameras.
Finally, time series plots (N
Sd
2
96.1 , t) are plotted in order to monitor the bias between
cameras over time.
4.7.2.2 Discrete checks
The purpose is to transfer absolute calibration coefficients from one camera to an other camera.
This may be needed when a significant bias between cameras is detected with continuous checks
on overlap zones and no (good) calibration imagery exist for one of the cameras.
One camera, which we assume to be well calibrated, is used as the reference camera. The
absolute calibration coefficients for this reference camera are denoted as k
refA . The calibration
coefficients for the camera which has to be calibrated are denoted as oldk
calA ,and
newk
calA ,. With
oldk
calA , the current calibration coefficient and
newk
calA ,the calibration coefficient to be retrieved
based on the overlap zone with the reference camera.
The DNs in overlap-intermediate-products (raw geometry) for both cameras are converted to
TOA radiances following Equation 4 using the current calibration coefficients.
kk
gik
refT
kkTgi
kgi,
kacquiredi,
kg
kacquiredi,k
refdITITgA
dITITdc off-)(DNNL -DNL
,,
,,
kk
gioldkcalT
kkTgi
kgi,
kacquiredi,
kg
kacquiredi,oldk
caldITITgA
dITITdc off-)(DNNL -DNL
,,,
,,,
The set of N (k
refL ,oldk
calL ,) pairs is used to fit a linear model through the origin (assuming correct
dark current calibration for both cameras):
k
ref
k
cal LbLModel )(
The least square estimators of b in the regression is obtained by minimizing
²1
,,
N
i
k
iref
k
ical bLLQ
This leads to the following estimator for b :
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 110/139
N
i
k
iref
N
i
k
iref
k
ical
L
LL
b
1
,
1
,,
)²(
The unbiased estimator of the error variance for the regression is
1
))(mod(1
,,
N
LelL
MSE
N
i
k
ical
k
ical
The 95% confidence interval for b is given by
N
i
k
irefL
MSEb
1
, ²
96.1
Following Equation 4 the new estimate for the calibration coefficient newk
calA ,is then calculated, as
b
AA
k
refnewk
cal ,
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 111/139
4.8 Multi-angular/Equalization
4.8.1 Introduction
The sensitivity of the detector response is not constant over the complete field of view. In order
to correct for the sensitivity variations, they first have to be accurately characterized. This is the
goal of multi-angular calibration and equalization.
The variations in detector response have two main causes:
1. variations in the sensitivity of individual pixel responses. This is a well known
phenomenon, typical for all solid state detectors.
2. variations over the field of view: the optical transmission which slightly decreases for
larger viewing angles;
They manifest themselves as different effects on different spatial frequencies. The concept is
illustrated in Figure 35. Therefore they can studied in separate components:
1. a high frequency component (HF)
2. a low frequency component (LF)
The LF corresponds to the variations in optical transmission, which can be described by a
polynomial. The HF only refers to the variations over individual pixels.
Figure 35: Sensitivity profile decoupled into high and low frequency components
The general approach to estimate the sensitivity variations is to image targets that are very
uniform so that observed variations reflect sensitivity variations rather than target variations.
Target uniformity is never perfect: the uniform reflectance of the targets is disturbed which can
be described as consisting of random noise and some structure. A measurement will always show
the non-uniformity of the targets combined with the sensitivity variations. This is depicted in
Figure 36: the same surface imaged by 2 different parts of a detector (with if non-uniformity
profile) yields different results.
To even out the non-uniformity, averaging over many measurements is performed on two levels.:
By summing over all pixels obtained sequentially in the along –track direction, we have within
target averaging. This decreases the noise of the target and the structural non-uniformity in the
along-track direction. However structural non-uniformity in the across track direction will persist
0,75
0,8
0,85
0,9
0,95
1
1,05
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
sensitivity profile
low freq comp (offset -0,1)
high freq comp (mean = 1)
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 112/139
despite the averaging. To alleviate this; averaging over multiple targets will be performed. The
structural target non-uniformity will then also be averaged out.
Figure 36: Surface profile, two detector sensitivity profiles and the combined results results
The best estimates for the different frequency contributions (LF, MF, HF) are obtained by a
combination of different methods. The LF can be reliably obtained using desert sites because for
those sites, accurate directional reference data can be obtained using BRDF models. The MF and
HF components will be estimated from measurements obtained over Antarctica.
4.8.2 Theoretical background
4.8.2.1 Sensitivity profiles
The sensitivity variations can be noted as: g(n) , where n represents the sensor pixel number. The
g(n) are normalized so that over the profile they have an average value of 1:
(n = 0,N-1) g(n) = N
For images acquired with the imaging system, any observed image line: f(n), is in fact the
product of the image line for a perfectly flat sensitivity profile f’(n) and the sensitivity profile
g(n):
f(n) = g(n) . f’(n)
f’(n) is assumed to contain no variations explicitly related to the pixel index n, only variations
coming from the image content.
4.8.2.2 Low and high frequency variations
The estimation of the complete function g(n) will be more difficult as it consists of different
contributions. Physically it is the result of several different causes: optical effects which cause
slow variations over the FOV and properties of individual pixels, which vary from one pixel to
the next. It is therefore justified to split up the sensitivity profile in to a low frequency and a high
frequency part:
0,55
0,6
0,65
0,7
0,75
0,8
0,85
0,9
0,95
1
19 24 29 34 39 44 49
surface
detector A
detector B
result A
result B
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 113/139
g(n) = g_low(n) . g_high(n)
The separation between the two can be made by a suitable frequency filter. If we define a low
pass filter LOW(). We can then divide the components as follows:
g_low(n) = LOW ( g(n)) g_high(n) = g(n) / g_low(n)
The filter should be chosen in a way as to facilitate the estimation of the separate parts, each
using their own techniques and images.
We can use a simple convolution filter in the spatial domain. Then the filter is as follows:
LOW ( g(n)) = g(n) (a,b)
in which (a,b) represents a boxcar function which equals 1/(b-a) inside the given interval,
and 0 outside of the interval. Such a filter, a convolution of the profile with a flat function of 17
pixels wide, was used in the experiments using Antarctica further on.
However, such a filter has some drawbacks. Therefore, it would be better to use a well designed
frequency space filter (as known from digital signal processing) such as the Hanning filter. This
will result in a more favourable splitting of high and low frequency components.
4.8.2.3 Estimating sensitivity profiles
The estimations of g(n) will be based on observed image lines. For the estimations, only the pixel
index n is of importance, so the estimation can be worked out one-dimensionally.
Since image content and sensitivity are inevitably combined, the g(n) cannot be retrieved from a
single line or a single image. However, if a collection of images and image lines is used, the
image content variations can be averaged out. The g(n) variations however are tied to pixels so
they do not average out.
For the estimation of the sensitivity profile g(n) we use a large collection of k image lines
(collected from a set of images):
{ f(n) }_k or { g(n) . f’(n)}_k
The difficulty is that the sensitivity variations g(n) can only be observed together with image
content variations f’(n)
A successful estimation of g(n) w will need to cancel out the effects of image content. It is
preferably based on a combination of two aspects:
avoiding image content variations by imaging homogeneous areas
averaging out remaining variations in a scheme in which many independent images are
combined.
Because of the different frequency components in g(n), it is best to perform its estimation using
two separate methods and separate imagery for the high frequency variations g_high(n) , and for
the low frequency variations g_low(n). Then, images can be used with properties which are
better suited for the estimation of one type. Finally the results of both estimations need to be
combined as to achieve one estimate for the complete sensitivity profile:
In the following we assume a perfect filter: Estimation of g_high(n) starts from:
f(n) = g_high(n) . g_low(n) . f’(n)
We take the low pass part of this:
LOW ( f(n) ) = LOW ( g_high(n) . g_low(n) . f’(n) )
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 114/139
For a perfect filter, this results in:
LOW ( f(n) ) = g_low(n) . LOW (f’(n) )
And the complementary high pass signal:
f(n) / LOW ( f(n) ) =
= ( g_high(n) . g_low(n) . f’(n) ) / g_low(n) . LOW (f’(n) )
= g_high(n) . f’(n) / LOW (f’(n) )
Thus the signal we intend to estimate is:
g_high(n) = ( f(n) /) . (LOW (f’(n) ) / LOW ( f(n) ))
or:
g_high(n) = ( f(n) / LOW ( f(n) )) ) . (LOW (f’(n) ) / f’(n) )
Now if the collection of images areas can be chosen in such a way that for after overall averaging,
f’(n) ≈ LOW (f’(n) ) This means that if the images either have to have no sifgnificant high frequency features or they
need to be canceled out by averaging. This is much easier to achieve than having all features
cancelling out.
If this is achieved, a good estimation of g_high(n) can easily obtained by averaging out the f(n) /
LOW ( f(n) )) over the available collection of image lines.
A similar estimation can be made for the low frequency components using images with very little
low frequency content.
4.8.2.4 Choice of images
The method explained above allows estimation of the high frequency part of a sensitivity profile
using suitable images. Ideal targets for this can be over the DOME-C site on Antarctica. The
extensive snow fields are contain only minimal high frequency content.
However, these images are completely unsuitable for the estimation of the low frequency
variations, as the images over Antarctica exhibit very strong BRDF effects. For the low
frequencies, we will therefore use different image types. Most suitable for this will be either
images of Deep convective clouds or Rayleigh scattering images
4.8.3 Algorithms
4.8.3.1 Sensor data collection
The multi-angular calibration will use data acquired from the desert sites for the determination of
the LF component. For the HF and MF components, data acquired over Antarctica will be used.
The calibration requires a good characterization of directional variation of the TOA reflectance.
During a period of about one month most, of the deserts sites will be viewed by all the detectors.
For each camera a polynomial function will be fit through the ratio of measured TOA reflectance
to modelled TOA reflectance of a given desert site as a function of viewing angle. This approach
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 115/139
has been applied by CNES for Polder (Hagolle et al., 1999 [LIT3]) and VGT. We will now
describe in detail the steps of the algorithm.
4.8.3.1.1 Data collection
The desert sites listed in Table 14Error! Reference source not found. are the areas of interest.
For each of them, the image data needs to be collected from the image segment, as digital
numbers as captured by the sensor for the different spectral bands.
Additionally, some ancillary data will be stored. This includes:
1. identification of the segment
2. applicable gain
3. absolute calibration coefficient.
4. average altitude
Some information about the atmospheric and meteorological conditions of the sites are also
needed. Therefore it is necessary to collect for (the centre of) every site following measurements:
5. water content
6. ozone content
Every desert is defined as a rectangular area described by the precise geographic coordinates of
the four corners. For every spectral band; the coordinates are converted to pixel coordinates on
the segment .In general they do not form a regular rectangle. We construct an enclosing
rectangle in pixel coordinates, as depicted in Figure 37 and store that for convenient access.
Figure 37: enclosing rectangle in image coordinates for a rectangle in geographic
coordinates
4.8.3.1.2 Data processing
1. Clean up data:
Unusable pixels need to be identified. A pixel is unusable if it is saturated or potentially cloudy.
If a site contains too many cloudy pixels, it is discarded completely.
2. Average Data:
For every detector (pixel along the line of the sensor) with coordinate falling into the enclosing
rectangle, calculate the line segment to average. These can be obtained by calculating the
intersection points between the lines that form the rectangle of the site, and the lines imaged by
a single detector. Different cases of such line segments are shown in Figure 38.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 116/139
Figure 38: line segments to average within the rectangle of interest
Once the line segments have been defined, the averaging is performed by simple summation over
the line segments (with x= i as line number, and y running from yi0 to yi1
1
),()(
1
01
i
io
y
yyii
i yiDNyy
DN
The average results are stored, together with the number of lines over which is averaged.
4.8.3.1.3 Radiance computation
The DN values can be converted radiance using the standard inverse sensor radiometric model
(cfr. section 4.2.2), using the following relation :
kk
gikT
kkTgi
kgi,
kacquiredi,
kg
kacquiredi,k
iTOAdITITgA
dITITdc off-)(DNNL -DNL
,
,,,
4.8.3.1.4 Conversion between radiance and TOA reflectances
The relation between the TOA reflectances kTOA,j and the radiances L
kTOA,j was explained in section
4.4.1.2.1.1. When all other variables are kept constant, both quantities are proportional. The
relation between the two can also be written as:
2
00 )cos(
1
d
dEL s
kk
TOA
k
TOA
The conversion depends on the solar zenith angle s for every detector. They are computed using
bilinear interpolation from the solar zenith angles provided in the ancillary data of the segment.
The TOA reflectance is corrected for gaseous transmittance as described under section 4.4.1.2.1.2.
4.8.3.2 Reference data computation
The principle behind the determination of the equalization coefficients is the comparison between
reflectances obtained from sensor measurements and modelled reference reflectances. The
computation of the reference radiances is explained in 4.4.4.2.1.
4.8.3.3 Computation of equalization coefficients
4.8.3.3.1 Ratio of in-flight versus pre-flight
Given the measured and modelled reflectances, we can compare them to derive equalization
coefficients. We recall that the DN is proportional to the LkTOA,, the precise relation (for the
nominal gain, ie. AVGg=1) is :
kgi
kkT
kkgi
kkTgi
kg
kTgacquiredj
gdITITALoffdITITdcNLDN ,,,,,,,.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 117/139
This relation holds with the absolute calibration coefficient Akpreflight and the equalization
coefficients gkj,g, pre-flight as determined before the launch.
In flight, we can calculate from the measured DN and the pre-flight coefficients, the Lk
jm, pre-flight and
the kj,g, preflight. On the other hand we have the modelled
kj,g, inflight values. We take the ratio
between the two:
k
flightinjgk
flightinkk
gikk
Tgikg
kTgacquiredj
kflightprejg
kflightpre
kkgi
kkTgi
kg
kTgacquiredj
kflightprejg
kflightinjg
kflightprejg
kflightinjg
gAdITIToffdITITdcNLDN
gAdITIToffdITITdcNLDN
L
L
,,,,,,,
,,,,,,,
,
,
,
,
)(
)(
Thus the ratio of the in-flight vs the pre-flight calibration coefficients is given by the ratio of the
reflectances:
kflightinjg
kflightprejg
kflightprejg
kflightpre
kflightinjg
kflightink
jggA
gAR
,
,
,
,
It is useful to plot the ratio Rkjm as a function of the angle.
4.8.3.3.2 Equalization profiles of the detectors
In the previous result the absolute calibration coefficient Ak and the variations at all different
frequencies are combined. The equalization coefficients gkjg contain these variations, which were
categorized before into HF, MF and LW components. We can split the equalization into a high,
medium and low frequency part:
gkjg = ghf
kjg gmf
kjg glf
kjg
If we assume that two of the three terms have been characterized using other means, we can
determine the remaining term.
4.8.3.3.3 Determination of the low frequency term : polynomial fit
The low frequency terms glkjg should model the slow variations due the optics. It is good practice
to model these variations with a function with a limited number of parameters. We propose to use
a fifth degree polynomial:
p
p
kpg
kjg
kjg jrRPolyR
5
0,
The coefficients rkg,p are optimised to minimize the quadratic difference between R
kjm and
RPolykjg. This polynomial regression is obtained by using a standard technique for general least
squares optimisation, as explained in (Press et al, 2007, [LIT45]). This results in a smoothly
varying function.
The residues, not modelled by the polynomial, show high frequency variations. However, these
can be disturbed by artefacts caused by the non-uniformities in the desert targets. Because of this;
they are not sufficiently accurate to be used for determining the gmkjg and gh
kjg coefficients.
Therefore, will will estimate these coefficients with other target, namely over Antarctica.
4.8.3.3.4 Outlier handling
A concern with least squares optimisation is that its results can be easily disturbed by outliers.
For the initial pre-flight determination of glkjg, special care must be taken to remove any outliers
by manual inspection.
The in-flight low frequency variations can be assumed to be at most slowly varying over time.
Therefore we can do the following:
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 118/139
- normalize the in-flight glkjg; by dividing them by the pre-flight or previous in-flight RPoly
kjg
- calculate the mean and standard deviation of the normalized values (averaged over all angles)
and determine a confidence interval around the mean
- all values outside of the confidence interval are assumed outliers and marked
- The new in-flight polynomial is calculated without using the outlier points
4.8.3.4 High and medium frequency (Antarctica)
The high and medium frequency terms can be assessed during flight using a statistical approach.
The idea is to average over enough measurements so that the high frequencies stemming from the
targets are effectively averaged out.
This can be best achieved with large, bright targets that have only limited high frequency content.
Extensive snow covered areas over Antarctica are well suited for this. The DOME-C site, already
presented in section 4.5.4.1.3 will used.
The image acquisition will be done as outlined in section 4.5.4. The area is not part of normal
operational imaging, so additional dedicated acquisitions with special settings need to be
performed for several orbits a day. This can only be done during local summer, which limits the
usability of the method.
The images are averaged out over the lines in the same way as in section 4.8.3.1.2. The result
contains variations at all frequencies. The low frequencies, which reflect both the target and the
imaging variations, need to be separated and removed. This is done by constructing a polynomial
model of the data, in the same way as proposed in 4.8.3.3.3 The data is normalized by the model
values, so that low frequency variations are removed.
The result still contains both high and and medium frequencies. They both need to be applied in
the process of multi-angular equalization. There is no explicit need to separate the components.
However, in order to analyse the results and their evolution over time, it is useful to separate the
terms .This can be done with standard signal processing techniques: the high frequency part can
be isolated well by applying a Hanning filter in frequency space, as proposed in (Fougnie et al
2000,[LIT46]).
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 119/139
4.9 Performance Indicators
As there is no on-board light source available on PROBA-V, an assessment of the instrument’s
image quality in flight will have to be done using ground targets. To determine instrumental noise
effects on the measured signals, ground targets should be spectrally uniform over the measured
range. As established elsewhere in this document, ground targets such as desert zones or snow
zones are considered as uniform zones. To determine image contrast, here done by measuring
instrument system MTF, ground targets should have features of high contrast.
4.9.1 Noise analysis
4.9.1.1 Background
When an image is captured, noise measurements are affected both by the noise effects caused by
the instrument capturing the image, as well as by variations in the image itself. To isolate the
influence of these image variations, one needs to perform the noise analysis over selected zones
which are spectrally uniform with respect to the instrument band considered in the analysis.
These uniform zones typically only comprise a sub-region of an entire image, for an instrument
with a low resolution and large swath such as PROBA-V. Therefore, to estimate noise over the
full FOV, multiple zones need to be selected. A noise analysis is complete only when a
sufficiently uniform zone is found and processed for each band, and this for the entire FOV.
Aside from image noise, several different noise influences can be analysed. These are discussed
in section 4.9.1.3. In section 4.9.1.2, the different statistical methods of analysis are described.
4.9.1.2 Noise analysis methods
Averaged pixel value: over different scan lines of the same uniform zone, a pixel value
should be more or less the same. Therefore, the averaged pixel value is a useful measure:
llpp L ,
Noise over pixel value: deviations of different scan lines with respect to the averaged
pixel value are a measure of the uniformity of the captured values, and therefore of the
noise:
lplpplplp LLRMSE 2
,, )()(
Averaged region value: over different scan lines of different pixels covering a uniform
region, all averaged pixel values should be more or less the same. Therefore, the average
over all pixels in that region, or over all non-overlapping zones covering that region, and
lines is a useful measure:
zz
pplplpregion L
,,
High Frequency (HF) noise over region value: a measure of the HF noise over an
entire region with respect to subdividing zones is the RMSE of the noise over zone of
pixel values covered in that region:
)(, zzregionHF RMSE
Low Frequency (LF) noise over region value: a measure of the LF noise over an entire
region with respect to subdividing zones is the RMSE of the deviations of averaged zone
values with respect to the averaged region value.
)(, regionzzregionLF RMSE
N-centred average: this determines the centred average (average including centre) for a
neighbourhood of N pixels:
2/
2/,
1
1Np
NpjjcenterNp
N
N-neighbourhood average: this determines the local average (average surrounding
centre) for a neighbourhood of N pixels:
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 120/139
2/
2/,
1Np
pjNpj
jlocalNp N
N-centred deviation: this determines the deviation of a averaged pixel value with respect
to the N-centred average value:
centerNppcenterNp ,,
N-neighbourhood deviation: this determines the deviation of a averaged pixel value
with respect to the N-neighbourhood average value.
localNpplocalNp ,,
Maximum N-neighbourhood deviation over region value : This determines the
maximum neighbourhood deviation over a region. This is a measure of the local noise:
)(max ,p, localNplocalNregion
Crucial remark: For N-centred and N-neighbourhood measures, a problem arises at the border
of an image, because fewer than N neighbours can be reached. Therefore, for border pixels the
local means are calculated over the useful pixels of the image, and N is reduced accordingly.
4.9.1.3 Noise influences
A noise analysis is performed for each spectral band separately. The following noise influences
need to be distinguished:
Image noise, or noise related to the ground target or object scene being measured. This
noise can be determined by averaging over the measured values of spectrally uniform
subsets of the image, and also calculating the noise terms over these subsets. The high-
frequency (HF) noise over the entire image then determines the noise induced by the
image. The low-frequency (LF) noise can also be calculated, this is a measure of the
uniformity of the image zone considered.
Inter-column noise, or deviations between pixels. This will include discontinuity effects
such as defective or degraded pixels. To examine this, first the averaged pixel values are
calculated, and then, the average and deviation of each averaged pixel with respect to its
nearest neighbours is calculated. The maximal deviation measured over pixels of the
entire array should be limited, in order for the image to be reliable. This measure can
therefore be used to mark suitable zones (i.e. zones where the maximal neighbourhood
deviation is within bounds).
Adjacent-array noise describes the noise effect caused by the merging of different
arrays to form one line of pixels. This noise will be found at the edges where array values
are merged in this manner. In PROBA-V, two effects need to be distinguished, which are
treated similarly. First, the butting of the SWIR arrays for each camera means that at each
end of two SWIR arrays, an adjacent-array noise effect results. Second, the overlap of
one camera to the next will result in another adjacent-array noise effect, and this for each
band. Such noise can be measured by calculating the mean time-averaged signal for a
neighbourhood of pixels at one edge in the first array, and calculating the mean time-
averaged signal for a neighbourhood of pixels at the adjacent edge in the succeeding
array. The difference between both is a measure of the noise between these arrays.
Equalization noise describes the differences caused by the equalization process, which is
performed for each uniform subset of the image separately. Hence, in a similar approach
to the inter-column noise calculation, one calculates the average and standard deviation
for each pixel with respect to its nearest neighbours. Instead of evaluating these local
averages over the entire array, the mean average and mean standard deviation are
calculated for each uniform subset separately. In this way, the equalization differences
between these uniform subsets are assessed.
Column noise, or noise related to a pixel of the detector. This noise will include both the
noise induced by the instrument (for each pixel separately), and the noise induced by the
image. By subtracting the image noise, the pixel noise or instrument column noise can be
determined.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 121/139
Structural noise can be detected by calculating the MTF of a time-averaged value of a
uniform subset of the image. An additional analysis can be performed by taking the two-
dimensional MTF of the entire uniform subset-image.
The assessment of image noise is not straightforward, as effects of atmospheric calculations and
physical atmospheric variations affect the reliability of the results. As an alternative method, the
calculated image noise values can also be extracted from data by other satellites (TBC).
4.9.1.4 Algorithm
4.9.1.4.1 Image noise
Input:
A set of image subsets (Level 1C), capturing scenes with zones which are spectrally
uniform over the mission’s band under test. Insofar as possible, the images of these zones
cover the entire FOV of the instrument.
status list of bad pixels
Procedure:
Eliminate all ‘bad’ pixels from the zone images using the bad pixel status list.
o See section 4.3.2.
Determine for every zone, the averaged zone value and noise value.
o For every zone z, determine lplpzregion L ,,
o For every zone z, determine HF noise
))(( ,, zlplpzregionHF LRMSERMSE
Eliminate all zones with significantly large HF noise.
o First, it can be defined by the user when a noise figure is considered too large.
o Next, all zones with HF noise larger than the user-defined limit are eliminated.
Determine the averaged image value and the image LF and HF noise:
o Determine zzregionimageregion .
o Determine )( ,, zregionHFzimageregionHF RMSE .
o Determine )(, imageregionzzimageregionLF RMSE .
Output:
The averaged image value is a measure of the signal level, and is used to identify whether
the test describes an L1, L2, L3 or L4 equivalent radiance.
The HF noise measure corresponds to the measured noise over the image.
The LF noise measures the uniformity of the considered zone.
4.9.1.4.2 Inter-column noise
Input:
A set of image subsets (Level 1C), capturing scenes with zones which are spectrally
uniform over the mission’s band under test. Insofar as possible, the images of these zones
cover the entire FOV of the instrument.
status list of bad pixels
Procedure:
Eliminate all bad pixels from the zone images using the bad pixel status list.
o See section 4.3.2.
Determine for every zone, the averaged pixel values and averaged zone value
o For every zone z, determine llpp L , .
o For every zone z, determine ppz .
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 122/139
Determine for every zone the instrumental noise by subtracting LF noise:
o For every zone z, determine
2/
2/,
1
1Np
NpjjcenterNp
N . This represents
the local image contribution.
o For every zone z, determine centerNppcenterNp ,, . This represent the local
instrument noise.
Determine for every instrumental noise, the N-neighbourhood average and
deviation:
o For every μp, determine
2/
2/, ,
1Np
pjNpj
localNp centerNjN
. Note: bad pixels are not
counted in value, but their index is used in the definition of the neighbourhood.
o For every μp, determine localNpcenterNplocalNp ,,,
Determine for every zone the maximum N-neighbourhood deviation, the local HF
noise and maximum difference between N-neighbourhood averages :
o For every zone z, determine )(max ,p, localNplocalNz . This is a measure for
the inter-column noise in this zone.
o For every zone z, determine )( ,,, localNpplocalNzHF RMSE .
o For every zone z, determine
zz NcenterNpNcenterNplocalNzLF ,1,N,, z
max , with Nz ranging
from beginning of the zone to end of the zone in steps of Nz pixels.
Determine the averaged image value, maximum maximorum N-neighbourhood
deviation, the maximum local HF noise, and the maximum maximorum difference
between N-neighbourhood averages:
o Determine zz nintercolum.
o Determine )(max ,zintmax, localNzercolumn .
o Determine )(max ,,localNn,intercolum, localNzHFzHF .
o Determine )(max ,,localNn,intercolum, localNzLFzLF .
Output:
The averaged value is a measure of the signal level.
The maximum maximorum N-neighbourhood deviation measures the maximum inter-
column noise present. This measure has a conservative measure (it over-estimates the
noise distributed over the considered image subsets). It can be called the maximum
intercolumn noise.
The maximum local HF noise measures the nominal intercolumn noise, as it measures
the rms distribution of the local deviations. It is always lower than the maximum
intercolumn noise.
The maximum maximorum difference between N-neighborhood averages is a measure of
the local LF noise; it determines the noise over a frequency of Nz pixels.
4.9.1.4.3 Adjacent-array noise
Input:
A set of image subsets (Level 1C), capturing scenes with zones which are spectrally
uniform over the mission’s band under test. The images of these zones are chosen to be
centred over the region between two adjacent arrays. This can either be the region
between two adjacent strips of the SWIR band, or the region between two camera’s for
all bands.
status list of bad pixels
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 123/139
Procedure:
Eliminate all bad pixels from the zone images using the bad pixel status list.
o See section 4.3.2.
Determine for every array, over its border region, the averaged pixel values, and the
averaged array value o For every couple of bordering arrays arr1 and arr2, a border region B can be
defined, where
MjjjNj
Mjj
jNj
ppppB
arrforppp
arrforppp
,,
,
,
1
21
1
o For every array arr, determine llpp L , .
o For every array arr, determine pparr .
Determine for every border region, the averaged border value and the LF border
deviation.
o For every border region B, determine 2
21 arrarr
B
.
o For every border region B, determine 21, arrarrBLF .
Determine the averaged image value, the maximum adjacent array noise and the
nominal adjacent array noise.
o Determine BBarrayadjacent .
o Determine )(max ,arrayadjacent max, BLFB .
o Determine )( ,arrayadjacent ,min BLFBalno RMSE .
Output
The averaged value is a measure of the signal level.
The maximum adjacent array noise is a conservative measure of the noise expected at a
adjacent array crossing.
The nominal adjacent array noise is a measure of the distribution of noise over different
adjacent arrays. The difference between the nominal and the maximum noise is a
measure of the discrepancy between different adjacent arrays.
4.9.1.4.4 Equalization noise
Input:
A set of image subsets (Level 1C), capturing scenes with zones which are spectrally
uniform over the mission’s band under test. Insofar as possible, the images of these zones
cover the entire FOV of the instrument.
status list of bad pixels
Procedure:
Eliminate all bad pixels from the zone images using the bad pixel status list o See section 4.3.2.
Determine for every zone the averaged pixel value over a fixed number (TBD 500)
of lines
o For every zone z, determine llpp L , .
Determine for every zone the instrumental noise by subtracting LF noise:
o For every zone z, determine
2/
2/,
1
1Np
NpjjcenterNp
N . This represents
the local image contribution.
o For every zone z, determine centerNppcenterNp ,, . This represent the local
instrument noise.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 124/139
Determine for every zone the averaged zone instrument noise, the HF zone noise
o For every zone z, determine pcenterNpzinstrument ,, .
o For every zone z, determine
)( ,,, zinstrumentcenterNppzHF RMSE .
Determine the average instrument noise, and the nominal image noise.
o Determine zzinstrumentinstrument , .
o Determine )( ,onequalisati nominal, zHFzRMSE .
Output:
The averaged instrument noise is a measure of the average local noise attributed to the
instrument, under the assumption that the image experiences no noise over a local
neighbourhood.
The HF zone noise is a measure of the distribution of the average local noise over that
zone. It gives a large value when the distribution is strongly fluctuating, which is an
indication of bad equalization.
The nominal image noise is a measure of the nominal equalization noise over the system.
4.9.1.4.5 Column noise
Input:
Image subsets used in the image noise calculation. Extract from them the raw values Lp,l ,
the averaged zone value zregion .
status list of bad pixels
Procedure:
Eliminate all ‘bad’ pixels from the zone images using the bad pixel status list.
o See section 4.3.2.
Determine for every zone the averaged pixel value over a fixed number (TBD 500)
of lines
o For every zone z, determine llpp L , .
Determine for every zone the instrumental noise by subtracting LF noise:
o For every zone z, determine
2/
2/,
1
1Np
NpjjcenterNp
N . This represents
the local image contribution.
o For every zone z, determine centerNppcenterNp ,, . This represent the local
instrument noise.
Determine for every zone, the global HF pixel noise value.
o For every zone z, determine global HF pixel noise
)( ,,, zlplzregionpHF LRMSE .
Determine for every zone the pixel column noise, as being the local HF contribution
not due to equalization noise. Determine the nominal zone column noise
o For every zone z, determine 2
,
2
,,,, centerNpzregionHFpzregioncolumnp .
o For every zone z, determine pzregioncolumnpcolumnz
2
,,, .
Determine for every zone the nominal image column noise.
o Determine zcolumnzcolumn 2
, .
Output:
The nominal image column noise is a measure of the column noise of the system.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 125/139
The nominal zone column noise is a measure of the column within the specified zone.
4.9.1.4.6 Structural noise
Input:
A set of image subsets (Level 1C), capturing scenes with zones which are spectrally
uniform over the mission’s band under test. Insofar as possible, the images of these zones
cover the entire FOV of the instrument.
Status list of bad pixels
Procedure:
Eliminate all ‘bad’ pixels from the zone images using the bad pixel status list: o See section 4.3.2.
.
Interpolate values for eliminated positions.
o Linear interpolation of values with adjacent pixel values at places where pixels
have been eliminated.
Determine for every zone, the 2D fast-fourier transform.
o For every zone z, determine )(2),(2 ,, lpfrlfrpyx LFFTfrfrFTyx .
Determine for every zone the averaged pixel value
o For every zone z, determine llpp L , .
Determine for every zone, the fast-fourier transform over the averaged pixel line
o For every zone z, determine )()( pfrpx xFFTfrFT .
Analyse curves o Mark high frequency content,
nyqyxcolumn
yx
xfrfrfrfrfor
frfrFT
frFT2.0,3.0
),(2
)(22
(TBC)
Output:
Signalization of structured noise when condition of high frequency content is fulfilled
4.9.2 MTF assessment
4.9.2.1 Background
A valid approach for MTF assessment is to compare MTF registration for two instruments,
imaging the same scene with similar spectral content, but of different spatial resolution. PROBA-
V images are of a fairly low resolution of 100 meter at best. Images of a higher resolution
satellite of 30 meter or better will therefore have a much better MTF at the spatial frequency
range of PROBA-V. The core assumption of this method is that the MTF of the higher resolution
instrument is close to perfect, so that the MTF of the image is assumed equal to the MTF of the
higher resolution instrument. The MTF of the lower resolution instrument of PROBA-V can thus
be derived from this. Since the projection and resampling algorithms of the PF cannot be
neglected as an MTF influence, the MTF estimation is performed on L2A images.
4.9.2.2 Algorithm
4.9.2.2.1 Image comparison technique
Step1: Select image couples
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 126/139
Select a N x N image of PROBA-V Level 2A together with a higher resolution image
over the same region. The higher resolution instrument should have similar spectral
response curves compared to PROBA-V to have a good comparison. A difference in
atmospheric conditions present when the images are captured also affects the results. This
can be avoided by using images acquired at about the same time (near-overpasses) and
selecting regions which experience stable atmospheric conditions.
Resample higher resolution image to fraction of PROBA-V resolution
Resample higher resolution image such that:
sLowyx
sHigh pixelsizeR
xR
pixelsize ReRe
11 , where Rx and Ry are integer values, with a
minimum value of 10 .
Match pixel and line shift of Higher resolution image by correlating the two images
(matching their features). Apply pixel and line shift such that both images are registered
in the same (p,l)-grid.
Apply 2D Fast-Fourier Transform to both images
Determine ),(2),(2 Re,Re lpLFFTfrfrFT sHighfrlfrpyxsHigh yx . The transform is
done on Rx*N x Ry*N points.
Determine ),(2),(2 Re,Re lpLFFTfrfrFT sLowfrlfrpyxsLow yx The transform is
done on N x N points.
Take the magnitude of these Fourier spectra and normalize by DC component
)0,0(2
),(2),(2
)0,0(2
),(2),(2
Re
Re
Re
Re
Re
Re
sLow
yxsLow
yxsLow
sHigh
yxsHigh
yxsHigh
FT
frfrFTfrfrMTF
FT
frfrFTfrfrMTF
The corresponding Fourier spectra need to be known at N x N frequencies. Points for
which the High Res spectrum is 0 are interpolated by taken the mean of its neighboring
values.
Determine Ratio of MTFs
Determine the ratio of image MTFs. This is a measure of the degradation of the PROBA-
V instrument wrt the higher resolution performance. This higher resolution performance
can be assumed to be quasi-ideal at the frequencies of PROBA-V
),(2
),(2),(2
Re
Re
yxsHigh
yxsLow
yxsystemfrfrMTF
frfrMTFfrfrMTF
4.9.3 SNR calculations
4.9.3.1 Background
The calculations of signal-to-noise-ratio (SNR) for each pixel is equivalent to the approach used
to calculate the column noise terms. The averaged pixel value can be taken as the signal per pixel.
This provides an estimation of the SNR due to the instrument alone.
4.9.3.2 Algorithm
Input:
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 127/139
Image subsets used in the column noise calculation. Extract from them the averaged pixel
values p , and the pixel column noises zregioncolumnp ,, .
Procedure:
Determine for every zone the SNR per pixel, defined as the ratio of signal over
noise:
o For every zone z, determine zregioncolumnp
p
pSNR
,,
.
Determine for every zone the minimum zone SNR, the average zone SNR and the
distribution on the zone SNR:
o For every zone z, determine )(minmin, ppz SNRSNR .
o For every zone z, determine ppSNRz SNR , .
o For every zone z, determine )( ,, SNRzppSNRz SNRRMSE .
Determine the guaranteed SNR, the average SNR and the SNR distribution.
o Determine )(min min,min zz SNRSNR .
o Determine zSNRzSNR , .
o Determine )( ,SNRzzSNR RMSE .
Output:
The guaranteed SNR is a measure for the minimal radiometric resolution of the
instrument.
o The average SNR and SNR distribution are useful as auxiliary specification of
the overall radiometric resolution of the instrument.
o The zone SNR specifications are similar measures, applied only to a specific
zone, and are also useful as auxiliary specifications.
The SNR per pixel is a useful measure of the radiometric resolution given for each pixel.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 128/139
4.10 Calibration Validation
In the previous sections the theoretical basis and processing steps of a number of PROBA-V
radiometric calibration methods are described. This section describes the validation approach of
the radiometric calibration using independent means like a comparison to other space-borne
sensors (described in section 4.10.1), ground-based measurements (described in section 4.4.2)
and APEX underflights (described in section 0).
The Committee on Earth Observing Satellites (CEOS) defines validation as the process of
assessing, by independent means, the quality of the data products derived from the system output.
One can distinguish 3 stages in validation (NASA’s validation hierarchy) :
Stage 1: Product accuracy has been estimated using a small number of independent
measurements obtained from selected locations at particular times. Some ground data
collection involved.
Stage 2: Product accuracy has been assessed over a widely distributed set of locations
and at a number of times via several ground data collection and validation efforts.
Stage 3: Product accuracy has been assessed and the uncertainties in the product well
established via independent measurements in a systematic and statistically robust way
representing global conditions.
As quantitative information will be deduced from PROBA-V images and temporal analysis is
crucial, a Stage 3 validation of the radiometric calibration is needed.
We distinguish pre-launch and post-launch validation activities:
Pre-launch validation activities include 1) the verification of radiometric calibration algorithms
and 2) the characterization of error budget.
During the pre-launch phase, the radiometric calibration algorithms will be tested and verified by
applying them to
VEGETATION archive data available at VITO. For radiometric calibration, before
launch, a number of SPOT-VGT scenes over Antarctica, deserts, sun glint, deep convective
clouds areas and Rayleigh sites will be processed using the PROBA-V algorithms. The
retrieved calibration parameters will be compared to the values in the calibration parameter
files. The aim is to validate the core of the algorithms including the RTF calculations with
6SV and/or Libradtran, assumptions made on surface properties (e.g.desert BRDF model and
parameters), etc. using real satellite data from a spectrally similar sensor. These algorithm
cores can be considered as ‘generic’ for spectrally similar sensors.
simulated PROBA-V images generated by the System Performance Simulator: For
radiometric calibration, before launch, simulated PROBA-V images generated by the System
Performance Simulator will be used as a supporting verification strategy. The aim is:
to evaluate and quantify the sensitivity of the vicarious calibration algorithms to specific
PROBA-V radiometric uncertainty sources (e.g. noise, compression, temperature
dependencies ...)
to validate those radiometric IQC algorithms for which no comparable SPOT-VGT data
exist (e.g. camera-to-camera calibration, linearity checks)
to validate that the IQC workflow can process images from PF (via the Central
DataBase). The SPS will generate images in a format that can be converted into a L1A or
L1B. This can be registered as input data in the central data base which can be read by
IQC.
For the different radiometric calibration methods described and prototyped, the error budget is
determined following this approach: the prototypes are run taking into account the uncertainty of
a number of input parameters (e.g. aerosol model). The output of the prototypes is analysed
statistically to determine the error budget associated with this uncertain input parameter.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 129/139
Post-launch validation of radiometric calibration methodology includes 1) refinement of
algorithms and 2) uncertainty evaluation. During the post-launch phase the radiometric
calibration algorithms will be tested and refined with PROBA-V data. A consistency analysis for
the various algorithms will be performed with PROBA-V and a cross-sensor comparison will be
made. Furthermore ground-based measurements and APEX underflights will be performed at an
ad hoc basis to validate the radiometric calibration coefficients by independent means. The
radiometric calibration algorithms often are based on radiative transfer code which necessitates a
number of input parameters that are uncertain. In case of reflectance-based methods, in-situ
reflectance measurements of homogeneous reference sites acquired simultaneously with the
satellite overpass also introduce errors and the accuracy of this reflectance-based method is
smaller than the radiance-based methods (Fox, 2004 [LIT47]). By applying the radiometric
calibration methods systematic or random errors are introduced. Random errorsare minimized by
averaging a large number of measurements. During validation systematic errors in one or more
methods introduced by e.g. target BRDF or aerosol load can be detected by consistency analysis
of the various absolute and relative radiometric calibration methods and by comparing multiple
sensors (cross-sensors).
4.10.1 Cross-sensor calibration over stable deserts
4.10.1.1 Introduction
Cross sensor calibration is the term used for describing calibration methodology which employs
data from sensors from other satellite systems, which is called "exogenous data".
Because the other systems can be assumed to be calibrated well, this is a valuable addition to the
methods acting on the data of the satellite itself. Also, it is of great interest to obtain calibrations
that are in line with those of other systems, especially with well established systems. For
PROBA-V this is even more important as the prime goal of the mission is to provide proper data
continuity and consistency with VGT. This makes proper cross sensor calibration between VGT
and PROBA-V essential.
The most straightforward way to perform cross-sensor calibration would be to use simultaneous
observations of the same location using different sensors. This is often not realizable in practice,
therefore observations taken at different times have to be used. The different time implies
different conditions; therefore a methodology to map observations to a common standard is
needed. This can be borrowed from other calibration methods (multi-temporal calibration). It
requires the use of stable test sites such as deserts (or Antarctica).
If we want to compare observations form different sensors, in general we will have to take into
account following elements that cause differences:
Differences between the imaging instruments:
spectral responses of the several band can differ: this may require spectral
resampling/adjustment
instrument geometry, instrument resolution
Differences stemming from the acquisition: different timing:
Different atmospherical conditions
Difference in geometry: different sun angles causing different illumination of the surface
Not exactly the same surface area is imaged
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 130/139
Slight instability of the site.
We will now describe a methodology to bridge the gap between the PROBA-V and exogenous or
reference data. The procedure is schematically geven in Figure 39. It mainly consists of the
following steps : (1) creation of reference database (including cloud detection, atmospheric
correction and averaging), (2) preparation of the PROBA-V data (ie. Preprocessing, cloud
detection, atmospheric correction and averaging), (3) Finding Comparable observations, (4)
Spectral adjustment of reference reflectance spectra to the PROBA-V spectral bands and (5) Data
comparison to retrieve the cross-sensor calibration coefficient.
The different steps are described in detail in the following paragraphs.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 131/139
Figure 39: Cross sensor calibration overview
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 132/139
4.10.1.2 Algorithmic implementation
4.10.1.2.1 Creation of reference database
For the creation of the BOA reflectance database of the reference sensor (ie SPOT-VGT1 or
SPOT-VGT2) the procedures (steps b to e) described under section 4.5.3.2.1 needs to be
applied.
4.10.1.2.2 Preparation of PROBA-V data
Following the procedures described under section 4.5.3.2.1 (steps a to d) the average BOA
reflectance for each site (new
BOA ) is obtained from the DNs per pixel.
4.10.1.2.3 Finding Comparable observations
For every PROBA-V observation, we search through the datasets of available exogenous
observations (i.e. from SPOT-VGT) for comparable observation, similar to the procedure
described in section 4.5.3.2.2. .If several acquisitions from a certain sensor match the criteria, the
average is retained after outlier selected. This gives VGTmeask
gBOA
,,
, . In order to remove the outliers
in the reference observations the results of deserts absolute calibration is used. If the absolute
calibration differs more than 20 % from 1 the observation is not used in the mean calculation..
4.10.1.2.4 Spectral adjustment: Conversion from exogeneous to PROBA-V sensor characteristics
Starting from pairs of observations, the general idea is to process the reference sensor data in order
to generate reflectance data as would be obtained by the PROBA-V sensor. In case the PROBA-V
spectral response curves significantly differ from those of SPOT-VGT a spectral
resampling/adjustment of the BOA reflectance spectra will be needed to correct for differences in
spectral sensitivities between the reference sensor and the PROBA-V sensor.
For this we will use the Rahman-Pinty-Verstraete (RPV) BRDF model parameters in function of
wavelength for each site.
First, the theoretical desert spectrum in function of wavelength ( elmod) is calculated for site g
with the RPV BRDF model for the corresponding sun and viewing geometry and spectrally
resampled to the spectral response curves of each band of the reference sensor (i.e. SPOT VGT1
or SPOT VGT2; k
VGTS )
For each band k (= BLUE, RED, NIR, SWIR ) of reference sensor (VGT1 or 2)
Calculate
0
0
mod
mod),(
,
)(
dkS
dS
k
VGT
elk
VGT
elVGTk
gBOA
With elmod= RPV_BRDF(site g,
new
s ,new
v ,new
Ideally, the values elVGTk
gBOA
mod),(
, obtained using the model would be identical to the actual VGT
measured BOA reflectance values
VGTmeask
gBOA
,,
,
In practice this will not be the case.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 133/139
We now intend to calculate an adjusted desert spectrum )( ,
,
mod VGTmeas
gBOA
eladjusted for which they
are identical:
VGTmeask
gBOA
VGTmeas
gBOA
eladjustedk ,,
,
,
,
mod, )(
Once the adjusted spectrum is known, the responses for all spectral bands as they were obtained
from the PROBA-V sensor can be calculated in a straightforward way as follows:
For each band k (= BLUE, RED, NIR, SWIR ) of PROBA-V
Calculate
0
0
,
,
mod
mod),(
,
)(
)(
dkS
dS
k
VPROBA
VGTmeas
gBOA
eladjustedk
VPROBA
eladjustedVPROBAk
gBOA
The main issue is the calculation of the adjusted desert spectrum )( ,
,
mod VGTmeas
gBOA
eladjusted .
For each of the spectral bands, the solution has to obey the equality
VGTmeask
gBOA
VGTmeas
gBOA
eladjustedk ,,
,
,
,
mod, )( . This still leaves open a large number of spectra as
potential solutions. We will use a practical solution for which we determine an adjustment
spectrum )( ,
,
VGTmeas
gBOAAdj which is defined as:
elVGTmeas
gBOA
VGTmeas
gBOA
eladjustedk Adj mod,
,
,
,
mod, ).()(
As description of the model we choose a piecewise linear function, with a tie point at each of the
peak wavelengths (k
VGTpeak, ) of the VGT sensitivity curves. At each of the tie points, an
adjustment factor (k
VGTAdj ) will be determined. The adjustment spectrum is then described as
follows (assuming k
VGTpeak, are in ascending order)
)( ,
,
VGTmeas
gBOAAdj = (Adj ), ,
..: first
VGTpeak
lastfirstk
VGTAdj = first
VGTAdj
for first
VGTpeak,
= k
VGTk
VGTpeak
k
VGTpeak
k
VGTpeakk
VGTk
VGTpeak
k
VGTpeak
k
VGTpeakAdjAdj
)(
)(
)(
)(1
,,
1
,1
1
,,
,
for k
VGTpeak
k
VGTpeak ,
1
,
= last
VGTAdj
for last
VGTpeak,
For adjustment spectra given with this functional form, the calculation of VGTmeask
gBOA
,,
, can be split
into independent parts:
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 134/139
dS
dS
dS
dkS
lastVGTpeak
kVGTpeak
kVGTpeak
kVGTpeak
VGTmeas
gBOA
eladjustedk
VGT
VGTmeas
gBOA
eladjustedk
VGT
lastk
firstk
VGTmeas
gBOA
eladjustedk
VGT
k
VGT
VGTmeas
gBOA
eladjustedkVGTmeask
gBOA
,
,
1,
,
)(
)(
)(
.
)(
1)(
,
,
mod
,
,
mod
1
0
,
,
mod
0
,
,
mod,,,
,
For the first and last terms, the adjusted model equals the unadjusted one, except for the fixed
correction factors first
VGTAdj and last
VGTAdj . The middle terms can be split further into terms depending
resp and on1k
VGTAdj and
k
VGTAdj.
dSAdjdSAdj
dS
kVGTpeak
kVGTpeak
kVGTpeak
kVGTpeak
kVGTpeak
kVGTpeak
k
VGTpeak
k
VGTpeak
k
VGTpeakk
VGT
k
VGTk
VGTpeak
k
VGTpeak
k
VGTpeakk
VGT
k
VGT
VGTmeas
gBOA
eladjustedk
VGT
,
1,
,
1,
,
1,
)(
)(
)(
)(
)(
1
,,
1
,
1
,,
,1
,
,
mod
Writing out the complete equations shows that the adjustment factors k
VGTAdj appear all outside the
integrals. The integrals can be computed, reducing the equation to a linear one, containing the 4
variables k
VGTAdj . Such an equation is obtained for every band, so in total the problem is reduced
to a set of 4 linear equations
The set of equations can be easily solved using standard linear algebra techniques, resulting in
values for the adjustment factors, which cvompletley determined the adjustment spectrum, and in
turn the reflectance values for the adjusted model.
4.10.1.2.5 Data comparison
Finally, an estimate of the cross-sensor calibration coefficient is obtained as the ratio of new
BOA to
eladjustedVPROBA
BOA
mod),( .
eladjustedVPROBA
BOA
new
BOA
VPROBA
VGTPROBAVGTCross
A
AA
mod),(
_
/,
4.10.1.2.6 Outlier removal and daily averaging
A daily average is calculated by averaging the results obtained over the different sites after outlier
removal. A site is removed if it is detected as an outlier in at least one of the spectral bands. A
robust outlier selection procedure based on the median and standard deviation from the median is
used for this.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 135/139
4.10.1.3 Error sources
The radiometric inter-comparison focuses on the concordance of the radiometric calibration of the
different sensors. In order to be able to appreciate this concordance, it is necessary to know the order of
magnitude of external sources of uncertainties. These are:
- The differences of the spectral responses for the compared similar bands or ,if corrected
for, the uncertainty on the spectral adjustment factor
- The variation of the target reflectance itself due to the angular difference between two
compared observations or due to stability of the target
- The variation of aerosol (AOT and model)
- The inaccuracy of the geolocation of the viewed pixels for each sensors
- Uncertainty in meteo data (H2O, O3, P)
For the shorter wavelengths (especially blue), atmospheric uncertainty (aerosol and meteo) will have the
largest contribution to the total uncertainty. For the longer wavelengths, the variation of the target
reflectance will have the largest impact.
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 136/139
4.11 IQC-RC Straylight handling
4.11.1 Introduction
Straylight contamination can be present in both the nominal PROBA-V images and the specific
calibration images. The magnitude of the straylight contamination at a given pixel depends basically on 2
terms [PVDOC‐190]:
1. The attenuation of the input TOA radiance at the pixel by a factor (1-TIS) with TIS the
Total Integrated Scattering
2. The sum of the contribution of scattered light from all (neighbouring) pixels
Theoretically, a correction for the straylight contamination would be possible based on the PROBA-V
PSF’s for each band. However the uncertainty in the simulated PROBA-V PSF’s is almost 100 % (factor
1.7) (communication with PI) ; furthermore a straylight correction implementation has a significant
impact on the processing performance of the Processing Facility. Therefore the current baseline is not to
implement a straylight correction algorithm in the PROBA-V User Segment (i.e. PF and IQC). On the
other hand uncertainty in the IQC determined absolute calibration coefficient due to straylight
contamination should be reduced as far as possible. Here we describe the approaches that will be
followed in the IQC to limit straylight influence.
4.11.2 Impact on Calibration Methods and proposed actions
The largest impact is expected for those methods using high contrasting calibration scenes such as :
- Rayleigh calibration images over oceans with bright sources, mainly clouds in the
neighbourhoud.
- Deep Convective Clouds calibration images over oceans
- Sun glint Calibration images over oceans
- Lunar calibration : bright Moon against deep space
4.11.2.1 Rayleigh Calibration
The ratio Lcloud/Locean varies typically from around 10 in BLUE to 40 in RED and 80 in NIR. Therefore a
0.1 % straylight contribution relative to Lcloud, results in a 8 % straylight contamination relative to Locean in
the NIR. In the Rayleigh calibration method this NIR band is used as the reference band for the retrieval
of the aerosol optical thickness, which is in turn used to simulate with 6SV the radiance in BLUE and
RED bands. Straylight contamination has therefore both a direct and an indirect impact on the Rayleigh
retrieved calibration parameters for BLUE and RED band. In [PVDOC‐655] this impact has been
evaluated. It was concluded that even a 0.2 % straylight contamination (relative to Lcloud) would give an
additional 3.8 % error in the RED band calibration coefficients. For BLUE bands the effect is slightly
less (due to smaller Lcloud/Locean ratio) where a similar error is found for a 0.5 % straylight contamination
(relative to Lcloud).
The PROBA-V PSF gives strong straylight contamination near the source, however still some ‘diffuse’
straylight remains at larger distances from the clouds (roughly 10-9
for the NIR band at 2500 pixels from
the source [PVDOC‐190]. The PSFs give the contribution of a point source i.e. for only one pixel.
Therefore, although this number of 10-9
seems to be small at first sight, it becomes significant if large
parts of the image is covered by clouds, even if these clouds are situated far from the target pixel. As
currently the uncertainty in the simulated PSF’s is too high to calculate with sufficient accuracy the
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 137/139
maximum number of cloud pixels allowed in the image to keep the impact on the Rayleigh calibration
below a threshold, we propose the following approach :
- Only ocean pixels far enough from clouds will be kept in order to limit the effect of the
‘local’ straylight. This is performed by dilating the cloud mask with Ns pixels. During the
commissioning phase the optimal Ns value will be determined by image based analyses (i.e.
analyses of the pattern in the per-pixel retrieved calibration coefficient in function of the
distance to clouds). Dilation of clouds has been included in the standard Rayleigh workflow
as indicated in Figure 8.
- To analyse the effect of the ‘diffuse’ straylight the percentage clouds within a certain
region around each selected ocean pixel will be calculated (offline in a test-framework)
through a moving window technique. A 6312 x 1652 pixel window for the VNIR and a
3278 x 860 pixel window for the SWIR will be considered, similarly as the proposed
window in [PVDOC‐190]for straylight correction. Next, plots of the per-pixel retrieved
calibration coefficients against the cloud percentage for a series of images will be generated.
If a clear relationship is detected, an extra masking step based on percentage clouds has to
be added to the operational workflow.
- It might be that even with these two straylight analyses steps, i.e.‘distance to clouds’ and
‘percentage of clouds’ straylight contamination can not sufficiently be eliminated and that
the Rayleigh retrieved absolute calibration coefficients are therefore erroneous. This might
be revealed through comparison of the retrieved calibration coefficient obtained through the
different vicarious calibration methods. In case the Rayleigh calibration method gives
significant different results, it might finally be decided to no longer use this method for the
operational calibration of PROBA-V. For the Polder instrument for instance good
correspondence between the Clouds, Rayleigh, Deserts and Sun glint methods were found
except for the 443nm spectral band due to important and un-corrigible stray-light
contamination.
4.11.2.2 Deep Convective Clouds Calibration
The deep convective clouds method is an interband calibration method using the signal of specific
white clouds over oceans. The TOA radiance of these bright clouds can be affected by straylight
which gives an attenuation of the measured radiance. As long as this attenuation is the same in
BLUE, RED and NIR band, the interband calibration will be hardly affected by straylight. However
according to the simulations performed in [PVDOC‐190] this attenuation is band depended with the
strongest attenuation in the BLUE band. In this simulation the radiance in the BLUE band was
attenuated with roughly 3.8 % in the center of a 450 pixel wide circular input flux. For the RED band
this was roughly 2% and NIR band 1.5 %. Similarly the straylight simulations of a scene with one
side illuminated with L4 and the other side completely dark, gives respectively a 2 % , -1.1 % and -
0.9 % reduction of the L4 radiance at more then 2500 pixels from the edge. Although these numbers
might be quite uncertain and strongly depended on the location and size of the input flux, it gives a
clear indication that the straylight influence on Deep Convective Clouds pixels is probably band
depended. This has a direct impact on the accuracy of the method, especially for the BLUE band.
Furthermore, in [PVDOC‐190] a strong increase in the straylight (stronger attenuation) near the edges
of the bright source is visible, which is again band depended , varying roughly from -6% in BLUE, to
–3% in RED and -2% in NIR. Clearly if these strongly affected straylight pixels would be included
in the Deep Convective Clouds calibration method a 3% relative interband calibration accuracy is not
achievable for the BLUE band, taken into account the contribution of the other error sources as
discussed in Table 18.
As band-depended straylight attenuation of the bright signal of clouds will probably even be present
in the center of the clouds, an additional error due to straylight is probably un-evitable. In order to
reduce the impact the following action is proposed :
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 138/139
- Visual inspection of the per-pixel retrieved calibration coefficient to detect patterns
in function of the distance to the cloud edge. If any pattern is observed, a cloud
reduction step could be added to the workflow (currently not included) or
alternatively the size of the pixel window of the already included Homogeneity or
Cloudy Neighbourhood tests can be increased
4.11.2.3 Sun Glint Calibration
The Sun Glint Calibreation method is similarly as the Deep Convective Clouds method an interband
calibration method but which is also applicable to the SWIR band. As discussed in [PVDOC-611] the
specific design of PROBA-V with a band depended view geometry has already significantly reduced
the accuracy of the method, specifically for the SWIR band. A band depended straylight affect adds
an additional uncertainty to this.
In contrast to the Deep Convective Clouds a sun glint spot is relatively small in the across-track
direction, although it can be quite large in the along-track direction. Therefore it will be more
sensitive to straylight then the Deep Convective Clouds method. In order to assess the impact of
straylight on the Sun Glint Calbration the following approach will be followed during the
commissioning phase :
- Visual inspection of the per-pixel retrieved calibration coefficient to detect patterns
in function of the distance to the sunglint-water edge
4.11.2.4 Lunar Calibration
Straylight will increase the radiance in the deep-space region around the Moon and reduce if for
the Moon itself. This might cause errors in the delineation of the lunar disk, furthermore even when the
lunar disk delineation is done correctly, the measured disk resolved Lunar radiance will be reduced due to
straylight.
In case the instrumental straylight is not changing over time and if always the same error is made
in the lunar disk delineation, the impact on stability monitoring on the basis of the Moon might be limited
and therefore currently no action is foreseen. On the other hand if on a longer term the Moon would be
used for absolute calibration (when improved Lunar reflectance models come available) or for cross-
comparison with other missions, the straylight contamination would mean a serious problem and the
implementation of a straylight correction algorithm would be necessary
5. GEOMETRIC CALIBRATION
See [PVDOC-981] Algorithm Theoretical Baseline Document Geometric Calibration
PROBA-V Image Quality Center (IQC) ATBD (US-05) 22 Aug. 13
ref: 05.02_PV02/US/N77D7-PV02-US-5-IQC-ATBD-v1_8 139/139
Appendix 1 FLOWCHART CONVENTIONS
Flow concepts are shown diagrammatically throughout the document. The convention for the
various elements displayed in these diagrams is shown in Figure 40.
Figure 40: Conventions used in processing flow diagrams
Input/
Output
Data
Process
Decision End of loopStart of
loop